Technique for protecting an operation of a wireless access network

Abstract

As to one method aspect of the technique, which is performed by a wireless access network (500) comprising access points, APs (502), measurement values indicative of wireless links between a wireless device, WD (504; 1991; 1992; 2030), and multiple APs (502) distributed in the wireless access network (500) are obtained. The WD (504; 1991; 1992; 2030) is served (304) or initiated (304) to serve using at least one AP (502; 130) among the APs (502), wherein the at least one AP (502; 130) is determined using a central processor, CP (100), comprising a scrambler (110) and a machine learning system, ML system (103). The scrambler (110) applies the obtained (302) measurement values to an input layer (101) of the ML system (103) according to a permutation of the multiple APs (502), and an architecture of the ML system (103) is asymmetric with respect to the permutation of the multiple APs (502).

Description

[0001] Technique for protecting an operation of a wireless access network

[0002] Technical Field

[0003] The present disclosure relates to a technique for operating a wireless access network. More specifically, and without limitation, methods and devices are provided for mitigating an attack on a Distributed Multiple-Input Multiple-Output (D-MIMO) operation of the wireless access network.

[0004] Background

[0005] Modern and future wireless access network, such as radio or optical telecommunication networks, provide wireless connectivity to wireless devices (WDs) such as user equipments (UEs). The Third Generation Partnership Project (3GPP) is one of the international bodies for orchestrating the development and, through its members, the standardizing of the wireless access technology of such networks.

[0006] The increasing complexity of operating wireless access networks is successfully handled by machine learning (ML) systems that manage the resources of the wireless access network. The ML systems can "learn" the spatial correlation of operating access points (APs) distributed in the wireless access network based on ab initio physical simulations and measurements.

[0007] A prime example for such a correlated operation is a Distributed Multiple-Input Multiple-Output (D-MIMO) network, which comprises spatially distributed APs such as radio units (RUs), each comprising multiple antennas for MIMO transmission or reception. MIMO provides one or two degrees of freedom in forming (e.g., steering) a beam, one end of which is pinned to one AP. One technical feature of D-MIMO is full spatial diversity by moving the end of the beam and opportunistically forming virtual antenna networks for both transmission and reception. In particular, a decisive technical feature of D-MIMO is coherent joint transmission, which can achieve phase synchronization among APs. This means that the signals from different APs can be combined constructively at the WD, enhancing the signal strength and reducing the interference.

[0008] J. Yuan, et al., "Distributed FD-MIMO (D-FD-MIMO): from Concept to Field Test," 2022 IEEE Radio and Wireless Symposium (RWS), pp. 86-89, pointed out that UE- assisted over-the-air (OTA) calibration is the most important process of D-FD- MIMO.

[0009] D-MIMO networks are currently developed for future 3GPP Release 18 radio access networks (RANs) and beyond 5G (B5G) systems. A network controller, which is referred to a central processor (CP, or central processing unit, CPU), uses an Artificial Intelligence (Al) model, i.e. an ML system, to control key functions of the RAN such as power control and beam selection. Herein, in line with technical jargon, the terms CP and CPU may refer to network functions not related to, or not limited to, microprocessors as such.

[0010] Some of the present inventors pointed out in the scientific publication "Practical Adversarial Attacks Against Al-Driven Power Allocation in a Distributed MIMO Network", International Conference on Communications (ICC) 2023 in Communication and Information System Security Symposium (arXiv:2301.09305), that power control is crucial to optimize the spectral efficiencies of UEs and that max-min fairness (MMF) power control, which is a commonly used strategy, becomes intrinsically vulnerable against adversarial attacks when carefully crafted perturbations are applied to the input of the Al model. Detailed simulations revealed the effectiveness of adversarial attacks and the necessity of smart defense techniques.

[0011] Another critical function in D-MIMO networks, which is planned to involve Al models, is the process of beam selection. Beam selection includes AP selection and beam steering for uplink (UL) reception or downlink (DL) transmission. The Al model predicts the best beam (predicting both the best AP and the best beam steering) to serve a UE in the D-MIMO network based on a subset of beam measurements in the DL that have to be reported by UEs. Using the Al model enables to significantly reduce the size and complexity of this subset without restricting the predicted beam to said subset. While these functions are crucial to optimize the RAN and maximize spectral efficiency of the UEs served, existing Al models for handling such important functions, e.g. in D-MIMO networks, are vulnerable to malicious and in particular, adversarial attacks.

[0012] Specifically in existing Al-driven functions for beam selection in a D-MIMO network, an attacker can cause a dramatic decrease in network performance when malware, embedded in a compromised UE, targets the Al model for beam selection, which may be located in a CP of the RAN. With access to Reference- Signal Received Power (RSRP) values reported by the UE and the ability to train a surrogate Al model that imitates the actual Al model of the CP, the attacker can craft adversarial perturbations causing the Al model in the CP to select a suboptimal or incorrect beam. The impact of this kind of attack can be significant. The attack can degrade the communication signal strength by over 6 dB, which is a substantial loss. It is important to note that, due to the sequential nature of RSRP reports, the attacker has limited information at any one time, but the attack can still effectively mislead the Al model.

[0013] The problem is further aggravated because existing defense solutions against attacks, such as adversarial training, defensive distillation, and input denoising, have proven to be either ineffective in the face of adaptive adversaries or too resource-intensive for practical deployment.

[0014] Summary

[0015] Accordingly, there is a need for a technique that mitigates adversarial attacks without compromising the efficiency or effectiveness of Al-driven control of wireless access networks such as D-MIMO networks.

[0016] As to a first method aspect, a method performed by a wireless access network is provided. The wireless access network includes access points (APs) distributed throughout the network. The method involves obtaining measurement values indicative of wireless links between a wireless device (WD) and a respective one among multiple APs. The WD is served using at least one AP (e.g., exactly one AP) selected from among the APs of the wireless access network. The AP selection is carried out using a central processor (CP) that includes a scrambler and a machine learning system (ML system). The scrambler applies the obtained measurement values to an input layer of the ML system in accordance with a specified permutation of the multiple APs. The architecture of the ML system is asymmetric with respect to permuting the multiple APs at the input layer.

[0017] By apply the scrambling, embodiments of the method can obfuscate how the APs are input to the ML system, which architecture is asymmetric with respect to permuting the multiple APs at the input layer. Thus, the CP is capable of determining the at least one AP (or determining at least one of the wireless links or at least one RRM parameter value for the serving or at least one beam for the serving) according to its training using the same permutation, while the exact boundaries in the space of measurement values where the determination changes (i.e., where the output of the ML switches) depend on the permutation due to asymmetric architecture. Thus, the permutation establishes a secret, which makes it more challenging for an attacker to predict the relationship between the input and output of the ML system. In particular, the permutation can makes it almost impossible for an attacker to identify the boundaries in the space of measurement values where the determination changes, and thus can makes it very challenging to designing an effective perturbation to attack the CP, i.e. makes it very challenging to modify the measurement values to mislead the CP as an operational controller of the wireless access network.

[0018] Embodiments of the method can make an attack less effective or successful by virtue of the scrambling. Attackers, with the intent to mislead the ML system for AP selection (or beam selection or RRM parameter selection), will design their perturbation based on their understanding of the ML system and its sensitive points (e.g. the aforementioned boundaries). The attack aims at pushing the ML system beyond a sensitive point. Herein, sensitive points may refer to points (or any other subspace such as the boundaries) in the input space of the ML system where the output of the ML system switches (e.g., from one AP to another AP, or from one RRM parameter value to another RRM parameter value, or from one determined beam to another determined beam). However, when the scrambling is applied, these sensitive points change location in the solution space of the ML system due to its asymmetric architecture. This change is unknown to the attacker as the permutation is unknown. Thus, the perturbation designed by an attacker based on a pre-scrambling understanding of the ML system would no longer effectively exploit the sensitive points of the ML system. Consequently, if an attack is attempted, it is either ineffective because the sensitive points are missed or it can be directly detected at the CP because the perturbation is greater than an impact of any real-world propagation obstacle.

[0019] The method may be performed by one or more nodes of the wireless access network, e.g. one or more of the APs and / or a dedicated control node, e.g. the CP. The dedicated control node may be a node that does not provide wireless access or is not associated with a particular coverage area of the wireless access network or that does not include a physical layer or a medium access control layer or a radio link control layer of a wireless communication protocol stack. The dedicated control node or the CP may be a platform that hosts a RAN intelligent controller (RIC), which is responsible for controlling and optimizing RAN functions, e.g. according to in an open radio access network (O-RAN). For example, the dedicated control node or the CP may perform no baseband processing functions but higher- level functions such as network slicing, edge computing, and prioritized communications. Alternatively or in addition, the method may be performed by networked computing or decentralized processing, e.g. at an edge of the wireless access network (also referred to as edge computing) or spaced apart from the APs or in a central node such as a data center (also referred to as cloud computing).

[0020] The method may be implemented as a method of operating the wireless access network. Alternatively or in addition, the method may be implemented as a method of mitigating an attack on the wireless access network.

[0021] The machine learning system (ML system) may be referred to as ML model or artificial intelligence system (Al system) or Al model or Al agent or Al.

[0022] The measurement values may be based on bilateral signal exchanges between the WD and one of the multiple APs. In other words, each of the measurement values may be indicative of a WD link with uniquely one of the multiple APs. That is, every measurement value may correspond to a distinct wireless link from the WD to a specific AP out of the multiple APs. Alternatively or in addition, each of the multiple APs may be indicated by at least one of the obtained measurement values. In other words, each of the multiple APs may be characterized by a minimum of one obtained measurement value. The ML system may be trained to determine the AP based on the measurement values relating to the multiple APs. During the training, the measurement values may be applied according to the same permutation (i.e., using an equivalent scrambler).

[0023] The input layer of the ML system of the CP may be asymmetric with respect to the permutation of the multiple APs. Alternatively or in addition, the output layer of the ML system of the CP may be indicative of the at least one AP determined for serving the WD.

[0024] In other words, the first method aspect may be a method implemented by a wireless access network with numerous distributed APs. The method may comprise obtaining measurement values indicative of wireless links present between a WD and multiple APs. The method may further comprise a step of using a central processor (CP), which includes a scrambler and a machine learning system (ML system), to determine at least one AP from the (e.g., multiple) APs to serve the WD. In the CP, the scrambler applies the obtained measurement values to an input layer of the ML system as per a permutation of the multiple APs. The architecture of the ML system in the CP may be designed to be asymmetric in regard to AP permutations.

[0025] For example, the first method aspect may be a process operated by a wireless telecom network which includes access points (APs) such as base stations (BSs). The process comprises acquiring readings that refer to wireless channels (e.g., wireless connections) between a mobile station (MS, e.g. the WD) and several APs placed within the wireless telecom network. The process then comprises providing or starting to cater to the MS with a minimum of one AP out of the APs. This selection of AP is determined by a central processor (CP), which includes a cipher (e.g., the scrambler) and a system utilizing machine learning (ML system). The cipher applies the acquired readings to a first layer (e.g., the input layer) of the ML system as per an arrangement (e.g., the permutation) of the several APs. The ML system's structure (e.g., the architecture) is not symmetric with respect to a rearrangement of several APs.

[0026] The applying of the obtained measurement values to the input layer of the ML system according to the permutation by the scrambler may be referred to as scrambling, changing, permuting or shuffling of the multiple APs (e.g., of the identifiers of the multiple APs, and thus, the positions of the APs) before the ML system processes the measurement values. The applying of the obtained measurement values by the scrambler may comprise scrambling the order of the multiple APs, i.e. permuting the order of the multiple APs at an input of the ML system.

[0027] Each of the APs may be associated with and / or identified by an identifier (ID, e.g. an index or label), e.g. in the access network or in the D-MIMO site. The identifier may be fixed and / or unique (e.g., at least for the D-MIMO site of the D-MIMO operation or for the access network). The scrambling (e.g., the scrambling scheme) may comprise the permutation of the identifiers of the multiple APs.

[0028] The architecture of the ML system may refer to a connectivity within the ML system, e.g. the connectivity of layers of a neural network of the ML system. Examples of a layer of the neural network include a convolution or a filter. For example, the architecture of the ML system may refer to the connectivity of the input layer in the ML system. The architecture of the ML system may be s structure of the ML system that is independent of and / or underlying weights of a neural network or any parameter resulting from training the ML system.

[0029] Herein, the ML system performing a step (e.g., the step of determining, e.g. selecting) may further comprise that the ML system is trained to perform the step. For serving the WD according to the output of the ML system, the ML system may be trained to optimize the performance of the wireless access network, or enhance the user experience on the wireless access network by minimizing disruptions and maximizing overall network efficiency.

[0030] Serving the WD may comprise performing or initiating a wireless communication (e.g., a distributed MIMO communication or a multi-connectivity) between the wireless access network and the WD using the determined at least one AP.

[0031] The wireless link may comprise a network channel or a network connection, e.g. if the WD is in a connected state relative to the wireless access network. The obtained measurement values may be indicative of a link quality, a signal quality, and / or a signal strength of the wireless links. The WD may be wirelessly connected or connectable to the wireless access network. The WD may be referred to as user equipment (UE) or mobile station (MS). The WD may be a radio device. The wireless access network may be a radio access network (RAN). The WDs may be UEs according to 3GPP or MSs according to Wi-Fi. Alternatively or in addition, the wireless links may comprise free-space optical communication. The WD may access the wireless access network using optical resources of the wireless access network. Furthermore, any beam may encompass a directional electromagnetic wave (e.g., for reception or transmission), particularly in radio spectrum or optical spectrum (e.g., outside the visual spectrum).

[0032] From the perspective of the WD, the measurement values may be measured by receiving the wireless links (e.g., the reference signals on the beams) from the multiple APs. The WD may measure the measurement values (e.g., the link quality values such as RSRP values, RSRQ. values, and / or RSSIs) of the synchronization signals for radio resource management (RRM) and / or according to 3GPP document TS 38.533, version 17.7.0.

[0033] The determining (i.e., the inference phase) may be based on under-sampling of APs.

[0034] The ML system may determine the at least one AP among the multiple APs for which the measurement values are obtained. Alternatively, the ML system may determine the at least one AP among a set of the APs of the wireless access network, wherein the multiple APs are a proper subset of the set of APs. For example, the ML system may determine the at least one AP from among the multiple APs or from APs for which no measurement values were obtained. The subset of multiple APs for which at least one measurement values was obtained may also be referred to as the obtained APs or the reported APs. Similarly, the measurement values may be received or reported measurement values.

[0035] In an embodiment, the at least one AP may be determined from a set of APs of the wireless access network. The obtained measurement values may relate to the multiple APs which are a proper subset of the set of APs.

[0036] The set of the APs may comprise APs within a site of the wireless access network, e.g. a site of coherent service. The ML system may be trained to determine the AP (e.g., the best AP according to a utility function) out of the set of APs based on the measurement values relating to the subset of the multiple APs.

[0037] The determining of the wireless link for the serving may comprise ML-driven radio resource management.

[0038] In an embodiment, each of the obtained measurement values may be indicative of one of the wireless links using one AP of the multiple APs and one radio resource management (RRM) parameter value of an RRM parameter of the respective one AP.

[0039] The measurement values applied at the input layer of the ML system of the CP may be indicative of at least one wireless link for each of the multiple APs. Alternatively or in addition, the measurement values applied at the input layer of the ML system of the CP may be indicative of wireless links using at least one RRM parameter value for each of the multiple APs. Alternatively or in addition, the output layer of the ML model of the CP may be indicative of the at least one AP in combination with the RRM parameter value of the at least one AP determined for serving the WD.

[0040] Herein, RRM may encompasses system-level management of co-channel interference, radio resources, and / or other radio transmission characteristics in the wireless access network. A primary objective RRM may be to utilize limited wireless resources (e.g., temporal resources such as time slots, radio-frequency spectrum resources, and / or spatial resources such as beams) and / or limited infrastructure of the wireless access network as efficiently as possible. Alternatively or in addition, RRM may involve the CP for controlling the RRM parameter value of the RRM parameter such as transmit power, user allocation, beamforming, data rates, handover criteria, modulation scheme, error coding scheme, etc. For example, the ML system may be trained (in combination with the scrambler) to maximize a system spectral efficiency (e.g. measured in a bit / s / Hz / area unit or in Erlang / MHz / site), optionally while ensuring a level of user fairness between WDs accessing or connected to the wireless access network.

[0041] The WD may be served according to an operating parameter of the at least one AP. A value of the operating parameter may be determined using the CP. In an embodiment, the WD may be served according to an RRM parameter of the at least one AP. An RRM parameter value of the RRM parameter may be determined using the CP.

[0042] The output layer of the ML model of the CP may be indicative of the RRM parameter value determined for serving the WD. Optionally, the RRM parameter value determined by the CP for serving the WD using the at least one AP may be different from any RRM parameter value used for the one or more wireless links of the at least one AP.

[0043] The ML system may determine the RRM parameter value of the at least one AP. For example, an output layer of the ML system may output the determined RRM parameter value of the at least one AP. The ML system may select the at least one AP. In addition, the ML system may determine the RRM parameter value. For example, the ML system may be trained to determine the combination (i.e., the pair) of the at least one AP and the RRM parameter value.

[0044] The determining (i.e., the inference phase) may be based on under-sampling of RRM parameter values.

[0045] In an embodiment, the RRM parameter value may be determined from a set of RRM parameter values of the at least one AP. The one or more measurement values obtained for said at least one AP may relate to a proper subset of the set of RRM parameter values.

[0046] The combination of at least one AP and the RRM parameter value for said at least one AP may be determined from a set of combinations of AP and RRM parameter value of said AP, and wherein the obtained measurement values relate to a proper subset of the set of combinations. The set of RRM parameter values may be AP- specific (i.e., may be specific for the at least one AP or each AP). Alternatively, the set of RRM parameter values may be AP-independent. For example, the same set of RRM parameter values or the same number of RRM parameter values may be available for each of the multiple APs (i.e., each AP of the subset of APs, or even each AP of the set of APs). For example, the set of APs may comprise P APs and the set of RRM parameter values may comprise B RRM parameter values, so that the set of combinations comprises P-B pairs of AP and RRM parameter value. Furthermore, the obtained measurement values may be indicative of (i.e., may relate to) wireless links using a proper subset of the set of combinations.

[0047] In any embodiment, the RRM parameter value may be a wireless access capability of the respective AP. Herein, the respective AP may relate to the at least one AP or any one of the set of APs (e.g., in the context of determining the AP by means of the CP) or may relate to any one of the multiple APs (e.g., in the context of obtaining the measurement values).

[0048] In an embodiment, the RRM parameter may comprise at least one of transmit power, user allocation, beamforming, data rate, handover criterion, modulation scheme, and error coding scheme. Alternatively or in addition, the RRM parameter value may be indicative of at least one of a transmit power level, wireless communication resources allocated to the WD, a beam at the at least one AP, a data rate, a handover criterion, a modulation scheme, and an error coding scheme.

[0049] The determining of the RRM parameter value may comprise transmit power control, e.g. adjusting the value of the transmit power of the WD or the respective AP, optionally to optimize signal quality and / or minimize interference.

[0050] Alternatively or in addition, the determining of the RRM parameter value may comprise user allocation, e.g. allocating the WD or WDs to different resources of the wireless access network, optionally to maximize network capacity. Alternatively or in addition, the determining of the RRM parameter value may comprise beamforming. For example, the RRM parameter may comprise beamforming weights (e.g. a precoder for a TX beam or a decoder for a RX beam). The using the RRM parameter value for a wireless link and / or determining of the RRM parameter value may comprise focusing a transmission and / or a reception of signals in a specific direction, optionally to improve signal quality and coverage. Alternatively or in addition, the RRM parameter value may comprise data rates. The determining of the RRM parameter value may comprise determining an appropriate data rates for the WD (or for different WDs) based on the measurement values (e.g., channel conditions) and / or quality of service (QoS) requirements. Alternatively or in addition, the RRM parameter may relate to a handover of the WD (e.g., in a cellular wireless access network) or mobility of the WD (e.g., in a cell-free wireless access network). Alternatively or in addition, the RRM parameter value may comprise (e.g., specify or select) a handover criterion. For example, the determining of the RRM parameter value may comprise defining a condition under which the WD is to be handed over from one AP (e.g., a base station) to another to maintain seamless connectivity. Alternatively or in addition, the determining of the RRM parameter may comprise a modulation scheme. The determining of the RRM parameter value may comprise selecting an appropriate modulation scheme based on the measurement values (e.g., channel conditions) and / or data rate requirements of the WD. Alternatively or in addition, the determining of the RRM parameter may comprise an error coding scheme. Using the RRM parameter values for the wireless links or the determining of the RRM parameter value may comprise applying error correction codes to protect transmitted data against errors introduced by the respective wireless link.

[0051] The ML system may be trained (in combination with the scrambler) to manage cochannel interference, optimize wireless resources (e.g., radio resources or optical resources), and / or ensure efficient communication in the wireless access network.

[0052] The set of RRM parameter values may comprise at least one of a set of transmit power levels, a set of wireless resources allocated to the WD, a set of beams, a set of data rates, a set of handover criteria, a set of modulation schemes, and a set of error coding schemes. Analogously, the subset of parameter values comprise at least one of a subset of transmit power levels, a subset of wireless resources allocated to the WD, a subset of beams, a subset of data rates, a subset of handover criteria, a subset of modulation schemes, and a subset of error coding schemes.

[0053] The determining of the wireless link for the serving may comprise beam determination (e.g., beam selection).

[0054] In an embodiment, the wireless links may correspond to beams at the multiple APs. Alternatively or in addition, each of the obtained measurement values may be indicative of one of the wireless links using one AP of the multiple APs and one beam of the respective one AP.

[0055] In an embodiment, the WD may be served using beamforming at the at least one AP. A beam of the at least one AP may be determined using the CP.

[0056] The beamforming at the at least one AP may encompass forming the beam for reception and / or forming the beam for transmission. The serving of the WD may comprise determining at least one beam based on the obtained measurement values of a link quality of the beams using the ML system.

[0057] The determined beam may carry a data link, e.g. in the uplink or in the downlink of the wireless communication of the WD and the wireless access network.

[0058] In an embodiment, the beam may be determined from a set of beams of the at least one AP. The one or more measurement values obtained for the at least one AP may relate to a proper subset of the set of beams.

[0059] The determined beam may be selected or (e.g., linearly) combined from a set of the beams from the multiple APs. Optionally, the measurement values are indicative of a link quality of a proper subset of the set of beams.

[0060] The set of beams may correspond to the beams available (e.g., configurable) for the D-MIMO operation of the access network at the D-MIMO site or for the DC (or MC) of the WD.

[0061] The ML model may be trained to determine the beam out of the set of beams based on the link quality reported for only the proper subset of the set of beams.

[0062] Alternatively or in addition, the ML model may be trained to determine the beam as a (e.g., linear) combination of two or more beams in the set of beams. The two or more beams combined for the determined beam may originate from different APs.

[0063] In an embodiment, the ML system may determine a beam for the serving of the WD. Alternatively or in addition, the ML system may determine a combination of the at least one AP and a beam for the serving of the WD. Alternatively or in addition, the ML system may determine a downlink transmit power for the serving of the WD.

[0064] Following the determination of the AP and / or the beam and / or the RRM parameter value at the wireless access network, the wireless access network may configure the WD with at least one of an uplink transmit power and a precoder for an uplink beam. The ML system may determine at least one of the uplink transmit power and the precoder. In an embodiment, the measurement values may be obtained from the multiple APs. Alternatively or in addition, the measurement values may be indicative of the wireless links measured at the multiple APs with respect to the WD. Alternatively or in addition, the measurement values may be based on uplink reference signals received at the multiple APs from the WD.

[0065] The measurement values may be obtained by receiving one or more measurement reports from the APs, e.g. in a fronthaul or backhaul control signaling. The measurement values may be based on uplink measurements. Examples of the uplink reference signals include a sounding reference signal (SRS) and a random access preamble (e.g., a Zadoff-Chu sequence).

[0066] In an embodiment, the measurement values may be obtained from the WD. Alternatively or in addition, the measurement values may be indicative of the wireless links measured at the WD with respect to the multiple APs. Alternatively or in addition, the measurement values may be based on downlink reference signals received from the multiple APs at the WD. Alternatively or in addition, the obtaining of the measurement values may comprise receiving measurement reports comprising the measurement values, each of the measurement reports relating exclusively to one of the multiple APs.

[0067] The measurement values may be obtained by receiving measurement reports from the WD, e.g. in an uplink control signaling. The measurement values may be based on downlink measurements. The downlink reference signals may comprise at least one of synchronization signals (e.g., PSS or SSS), positioning reference signals (PRS), cell-specific reference signals (CRS), demodulation reference signals (DM- RS), phase tracking reference signals (PT-RS), and channel state information reference signals (CSI-RS).

[0068] In any embodiment, the WD may receive (downlink) reference signals from or transmit (uplink) reference signals to the multiple APs, i.e. some of the APs of the wireless access network for measurements resulting in the measurement values.

[0069] In an embodiment, the at least one AP may comprise two or more APs determined for serving the WD in dual connectivity (DC) or multi-connectivity (MC). The WD may be served using one AP from among the APs of the wireless access network. The CP may determine the one AP from among the APs of the wireless access network. Alternatively, the WD may be served using two or more APs from among the APs of the wireless access network, and / or the CP may determine the two or more APs from among the APs of the wireless access network for DC or MC.

[0070] In any embodiment, the operation of the APs for serving the WD and / or other WDs may cause a correlation in the determination of the at least one AP among the set of APs and / or in the determination of the RRM parameter value and / or in the determination of the beam.

[0071] In an embodiment, the multiple APs or the set of APs may be distributed across a site of coherent coverage of the wireless access network.

[0072] The coherent coverage may comprise a coherent joint transmission of the at least one AP to the WD and / or a coherent joint reception of the WD at least one AP.

[0073] In an embodiment, the wireless access network may serve the WD according to a distributed multiple-input multiple-output (D-MIMO) operation of the determined at least one AP or the set of APs. Alternatively or in addition, the multiple APs or the set of APs may belong to a D-MIMO site of the wireless access network. Alternatively or in addition, the obtained measurement values may be indicative of a link quality of at least one beam from each of the multiple APs belonging to a D- MIMO site of a D-MIMO operation of the wireless access network.

[0074] The ML model may be trained to determine the beam if the link quality is reported for at least one beam from each of the plurality of APs, e.g. each of the APs belonging to the D-MIMO site of the D-MIMO operation of the access network.

[0075] The ML system may be trained for the AP determination (e.g., AP selection) and / or beam determination (e.g., beam selection) in a D-MIMO operation of the wireless access network.

[0076] In any embodiment, by training the ML system in combination with the scrambler, the CP may be configured to determine the at least one AP that optimizes the coherent coverage and / or the D-MIMO operation taking the correlation between the APs into account. In an embodiment, the measurement values may be indicative of a link quality of each of the wireless links. Alternatively or in addition, the measurement values may comprise values of a reference signal received power (RSRP) or values of a reference signal received quality (RSRQ) or received signal strength indicators (RSSIs).

[0077] In an embodiment, the measurement values of the wireless links may be based on reference signals transmitted from the multiple APs on the respective wireless links. Alternatively or in addition, the reference signals may comprise synchronization signals periodically broadcasted in the downlink on the beams from the plurality of APs and / or primary synchronization signals (PSSs) and / or secondary synchronization signals (SSSs) and / or demodulation reference signals (DM-RS).

[0078] For example, the measurement values may comprise link quality values for the link quality of the beams from the multiple APs.

[0079] In an embodiment, an order of the multiple APs may be scrambled according to the same permutation during both training of the ML system and the determining of the at least one AP or the RRM parameter value or the beam using the ML system.

[0080] The same permutation (i.e., scrambling scheme) may be applied in both the training of the ML system and the determining step (e.g., the determining of the AP or the beam) using the ML system.

[0081] The scrambling scheme may comprise a one-to-one mapping (i.e., the permutation) of the multiple APs. For example, the multiple APs (e.g., the subset of APs) may comprise all APs of the wireless access network that belong to (e.g., that participate in or that are involved in) the D-MIMO operation at the D-MIMO site. The wireless access network may comprise one or more D-MIMO sites. The multiple APs (or the subset of APs) may relate to any one of the D-MIMO sites within the wireless access network.

[0082] The method may comprise the step of training the ML system (e.g. for determining the at least one AP, the RRM parameter value, and / or the beam). Alternatively or in addition, the method may comprise the step of determining the at least one AP, the RRM parameter value, and / or the beam. The training of the ML system may be referred to as the training phase. Alternatively or in addition, the determining (of the at least one AP, the RRM parameter value, and / or the beam) using the ML system may be referred to as inference phase.

[0083] The step of determining the at least one AP, the RRM parameter value, and / or the beam may be referred to as AP determination, beam determination, and / or beam determination (or beam selection), respectively. The determining step may be collectively referred to as determining at least one of the wireless links. The determined AP may be selected from the (e.g., discrete and / or finite) set of APs. The determined RRM parameter value may be selected from a (e.g., discrete and / or finite) set of RRM parameter values. The determined beam may be selected from a (e.g., discrete and / or finite) set of beams.

[0084] In an embodiment, the WD may be wirelessly connected to the wireless access network. Alternatively or in addition, the measurement values may be measured at the radio device, optionally measured for wireless access to the wireless access network or measured for mobility of the WD in a D-MIMO site of the wireless access network or measured during an idle or inactive state of the WD for early reporting enabling faster setup of the D-MIMO operation for the WD.

[0085] The WD may perform the measurements for the measurement values (e.g., the link quality) of the beams while the radio device is mobile at the D-MIMO site of the D-MIMO operation.

[0086] In an embodiment, the method may be performed by one of the multiple APs or one of the set of APs of the wireless access network. Alternatively or in addition, the method may be performed by the CP of the wireless access network. The CP may be hosting the ML system and / or the CP may be controlling the D-MIMO operation and / or the CP may be connected to each of the multiple APs or each of the set of APs.

[0087] The input layer of the ML system may encode, for each of the multiple APs, the measurement values of one or more wireless links (e.g., one or more RRM parameter values or one or more beams). Using the ML system for the determining (e.g., of the at least one AP or the beam) may comprise applying the obtained measurement values of the wireless links to the input layer. In an embodiment, the applying of the obtained measurement values may comprise permuting the order of the multiple APs according to the permutation of the scrambler at the input layer of the ML system without changing an order of RRM parameter values or beams within a feature group of the respective AP at the input layer of the ML system.

[0088] For example, an input plane of the ML model may be configured to receive, for each of the multiple APs, the measurement values (e.g., indicative of the link quality) of multiple beams or multiple RRM parameter values. Alternatively or in addition, the ML system may comprise multiple input planes. Each input plane may be sized according to the multiple APs (e.g., the subset of APs). Different input planes may be configured to receive different measurement values for the multiple beams and / or the multiple RRM parameter values relating to the respective AP.

[0089] Optionally, those of the multiples beams and / or the multiple RRM parameter values, for which no measurement value is obtained, may be left undefined or may be set to zero at the input of the ML system.

[0090] In an embodiment, the wireless access network may comprise multiple D-MIMO sites each comprising a CP and multiple APs for serving a WD in a D-MIMO operation. Different permutations of the multiple APs may be used for the applying of the measurement values obtained at the different D-MIMO sites.

[0091] In an embodiment, the permutations may be generated independently at each of the D-MIMO sites based on a source of true randomness other than a pseudorandom number generator and / or. The permutations of the D-MIMO sites may be mutually compared by exchanging between the CPs results of a hash function or a public key generator applied to an identifier of each of the permutations and / or. At least one of the permutations may be regenerated if the generated permutations are equal.

[0092] Embodiments can compare the permutations without disclosing the permutations outside the respective D-MIMO site (e.g., disclosing the permutations outside the respective CP), because a public key or a hash value is derived from the permutation (e.g., represented by an integer label corresponding to the permutation) using a one-way function such as the public key generator or the hash function, which means that it is computationally infeasible to derive the permutation from the exchanged result of the one-way function. Therefore, comparing the results of the one-way function does not reveal the underlying permutation. Each permutation may be assigned a unique integer label based on a position of the permutation in a lexicographically sorted list of all possible permutations.

[0093] In an embodiment, the method may further comprise receiving a configuration message indicative of the permutation. The configuration message may be encrypted.

[0094] While the scrambler has been described herein to apply (i.e., to map) the measurement values to the ML system according to a scrambling scheme comprising the permutation of the multiple APs, the scrambling scheme may further comprise replacing and / or omitting at least one of the multiple APs.

[0095] In an embodiment, the architecture of the ML system may comprise a convolutional filter or a convolutional neural network. Alternatively or in addition, the input layer of the ML system may comprise a convolutional filter or a convolutional neural network. Alternatively or in addition, a size of a kernel of the convolutional filter or the convolutional neural network or the input layer of the ML system may be multiple times smaller than the number of the multiple APs.

[0096] The convolutional filter or the convolutional neural network or the size of the kernel may cause an asymmetry with respect to permuting the multiple APs. Alternatively or in addition, the convolutional filter or the convolutional neural network or the size of the kernel may cause a local symmetry in terms of few APs. Alternatively or in addition, a robust scrambler may correspond to a permutation that maps neighboring APs to neighboring AP input features of the ML system. Alternatively or in addition, the permutation of the multiple APs may cause a global change of the order of AP input features of the ML system, e.g. which is not covered by the local symmetry.

[0097] As to a second method aspect, a method performed by a core node and / or a method of configuring a wireless access network comprising access points (APs) is provided. The method comprises a step of determining a permutation of multiple APs distributed in the wireless access network. The permutation is applied prior to a machine learning system (ML system) that processes measurement values indicative of wireless links between a wireless device (WD) and the multiple APs. An architecture of the ML system is asymmetric with respect to the permutation of the multiple APs. The method further comprises a step of sending a configuration message indicative of the determined permutation to a central processor (CP). The CP is configured to determine at least one AP. A scrambler of the CP applies the measurement values to an input layer of the ML system of the CP according to the permutation.

[0098] The second method aspect may be performed by any node of a core network supporting the wireless access network.

[0099] The second method aspect may be implemented as a method of deploying a wireless access network or controlling a wireless access network or a method of configuring a CP (or the scrambler of the CP) of the wireless access network.

[0100] The second method aspect may be implemented alone (e.g., by a core network) or in combination with any one of the embodiments of the first method aspect. The second method aspect may further comprise any feature and / or any step disclosed in the context of the first method aspect, or a feature and / or step corresponding thereto, e.g., a receiver counterpart to a transmitter feature or step.

[0101] While configuring the CP (e.g., the scrambler of the CP) has been described herein in terms of the permutation performed by the scrambler when apply the measurement values to the ML system according to a scrambling scheme comprising the permutation of the multiple APs, the configuration message may further be indicative of the scrambling scheme comprise replacing and / or omitting at least one of the multiple APs.

[0102] The determining of the permutation may also be referred to as determining the scrambler scheme or determining a scrambler.

[0103] In an embodiment, the determining of the permutation may comprise duplicating or replicating the ML system of the CP at the core node. In an embodiment, the permutation is determined dependent on the measurement values, optionally RSRP may value, obtained from the wireless access network or the WD.

[0104] In an embodiment, the permutation may be determined by selecting the permutation from a plurality of candidate permutations by measuring a performance measure of the duplicated or replicated ML system in combination with each of the permutation candidates. The duplicated or replicated ML system may be retrained for each of the candidate permutations.

[0105] The permutation selected based on the performance measure may correspond to a robust scramblers.

[0106] In an embodiment, the configuration message sent to the CP may be further indicative of the ML system retrained for the determine permutation.

[0107] In an embodiment, the wireless access network may comprise at least two sites each comprising multiple APs and a CP. The permutation may be determined for each of the at least two sites. Alternatively or in addition, configuration messages, each indicative of a respective one of the at least two permutations, may be sent to the respective CPs of the at least two sites.

[0108] The at least two sites may be different (e.g., spatially separated) D-MIMO sites. For example, the operation of the multiple APs within each site may be correlated and / or the operation of the multiple APs at different sites may be uncorrelated.

[0109] In an embodiment, different permutations may be determined for different sites of the wireless access network.

[0110] A plurality of different robust scramblers may be determined for the different sites.

[0111] In an embodiment, the sent configuration message or configuration messages may be encrypted.

[0112] While for clarity the technique is described referring to one WD (e.g., per D-MIMO site), the obtained measurement values may comprise measurement values from one or more further WDs camping on or connected to the wireless access network. The determined wireless link (i.e., the determined at least one AP, the determined RRM parameter value, and / or the determined beam) may prevent or reduce interference at the one or more further WDs when serving the WD.

[0113] Furthermore, while for brevity the technique is described to determine one RRM parameter value, the determining step may comprise determining multiple RRM parameter values using the CP for serving the WD according to the multiple RRM parameter values (e.g., simultaneously). The multiple RRM parameter values may comprise values for a combination of the RRM parameters disclosed herein. For example, different beams (according to the wireless links for the measurement values) may be combined (e.g., from the determined one AP or two or more determined APs) for the (e.g., coherent) joint transmission and / or joint reception.

[0114] Any WD may be a radio device. Any AP may be a base station. The one or more WDs (e.g. radio devices) and / or the wireless access network (e.g., a RAN) may form or may be part of, a wireless network (e.g., a radio network), e.g., according to the Third Generation Partnership Project (3GPP) or according to the standard family IEEE 802.11 (Wi-Fi). The first method aspect may be performed by one or more embodiments of the RAN or the CP or at least one of the APs. The second method aspect may be performed by one or more embodiments of a core node (e.g., a session management function, SMF) of the wireless network.

[0115] The RAN may comprise one or more base stations, e.g., performing the first method aspect. Alternatively or in addition, the radio network may be a vehicular, ad hoc and / or mesh network comprising two or more radio devices in pairwise direct radio connections such as sidelinks (SLs), e.g., acting as remote radio device and / or a relay radio device. The measurement values may be obtained through at least one SL.

[0116] Any of the WDs (e.g., radio devices) may be a 3GPP user equipment (UE) or a Wi-Fi station (STA). The WD may be a mobile or portable station, a device for machinetype communication (MTC), a device for narrowband Internet of Things (NB-loT) or a combination thereof. Examples for the UE and the mobile station include a mobile phone, a tablet computer and a self-driving vehicle. Examples for the portable station include a laptop computer and a television set. Examples for the MTC device or the NB-loT device include robots, sensors and / or actuators, e.g., in manufacturing, automotive communication and home automation. The MTC device or the NB-loT device may be implemented in a manufacturing plant, household appliances and consumer electronics.

[0117] Whenever referring to the RAN, the RAN may be implemented by one or more APs (e.g., network nodes, base stations, radio units or distributed units of base stations).

[0118] The WD may be a remote WD (e.g., remote radio device), which is wirelessly connected or connectable (e.g., according to a radio resource control, RRC, state or active mode) with a relay WD (e.g., relay radio device), which in turn is wirelessly connected or connectable (e.g., according to a radio resource control, RRC, state or active mode) with the at least one AP.

[0119] Any AP (e.g., network node or a base station) may encompass any station that is configured to provide wireless (e.g., radio or optical) access to any of the WDs (e.g., radio devices). The wireless access network may be a cell-free access network. Alternatively or in addition, any AP (e.g., base station) may comprise a cell, a transmission and reception point (TRP), a central unit (CU), a distributed unit (DU), a radio access node. The at least one AP (e.g. via the relay WD) may provide a data link to a host computer providing user data to the (e.g., remote) WD or gathering user data from the (e.g., remote) WD. Examples for the APs (e.g., base stations) may include a 3G base station or Node B (NB), 4G base station or eNodeB (eNB), a 5G base station or gNodeB (gNB), a Wi-Fi AP and a network controller (e.g., according to Bluetooth, ZigBee or Z-Wave).

[0120] The wireless access network (e.g., as a RAN) may be implemented according to the Global System for Mobile Communications (GSM), the Universal Mobile Telecommunications System (UMTS), 3GPP Long Term Evolution (LTE) and / or 3GPP New Radio (NR).

[0121] Any aspect of the technique may be implemented on a Physical Layer (PHY), a Medium Access Control (MAC) layer, a Radio Link Control (RLC) layer, a packet data convergence protocol (PDCP) layer, and / or a Radio Resource Control (RRC) layer of a protocol stack for the radio communication. Herein, referring to a protocol of a layer may also refer to the corresponding layer in the protocol stack. Vice versa, referring to a layer of the protocol stack may also refer to the corresponding protocol of the layer. Any protocol may be implemented by a corresponding method.

[0122] As to another aspect, a computer program product is provided. The computer program product comprises program code portions for performing any one of the steps of the first and / or second method aspect disclosed herein when the computer program product is executed by one or more computing devices. The computer program product may be stored on a computer-readable recording medium. The computer program product may also be provided for download, e.g., via the radio network, the RAN, the Internet and / or the host computer.

[0123] Alternatively, or in addition, the method may be encoded in a Field-Programmable Gate Array (FPGA) and / or an Application-Specific Integrated Circuit (ASIC), or the functionality may be provided for download by means of a hardware description language.

[0124] As to a first device aspect, a node of a wireless access network comprising access points (APs) is provided. The node comprises memory operable to store instructions and processing circuitry operable to execute the instructions, such that the node is operable to perform the first method aspect.

[0125] The node may further be operable to perform the steps of any embodiment of the first method aspect.

[0126] As to a further first device aspect, a node of a wireless access network comprising access points (APs) is provided. The node is configured to perform the first method aspect.

[0127] The node may further be configured to perform the steps of any embodiment of the first method aspect.

[0128] As to a second aspect, a core node for configuring a wireless access network comprising access points (APs) is provided. The core node comprises memory operable to store instructions and processing circuitry operable to execute the instructions, such that the core node is operable to perform the second method aspect. The core node may further be operable to perform the steps of any embodiment of the second method aspect.

[0129] As to another second device aspect, a core node for configuring a wireless access network comprising access points (APs) is provided. The core node is configured to perform the second method aspect.

[0130] The core node may further be configured to perform the steps of any embodiment of the second method aspect.

[0131] As to a still further aspect a communication system including a host computer is provided. The host computer comprises a processing circuitry configured to provide user data or process user data, e.g., included in user data exchanged with the WD in the serving step of the first method aspect. The host computer further comprises a communication interface configured to forward the user data to the wireless access network (e.g., a cellular network or the RAN) or the at least one AP (e.g., a base station) for transmission to the WD (e.g., a UE). A processing circuitry of the wireless access network (e.g., a cellular network) is configured to execute any one of the steps of the first and / or second method aspects. The UE comprises a radio interface and processing circuitry, which is configured to assist and participate in any one of the steps of the first and / or second method aspects.

[0132] The communication system may further include the WD (e.g., UE). Alternatively, or in addition, the wireless access network (e.g., a cellular network) may further include one or more APs (e.g., base stations) configured for wireless (e.g., radio) communication with the UE and / or to provide a data link (e.g., using the wireless link and / or AP and / or RRM parameter value determined by the first method aspect) between the UE and the host computer using the first and / or second method aspects.

[0133] The processing circuitry of the host computer may be configured to execute a host application, thereby providing the user data and / or any host computer functionality described herein. Alternatively, or in addition, the processing circuitry of the UE may be configured to execute a client application associated with the host application. Any one of the devices, the WD, the UE, the AP, the base station, the CP, the core node, the SMF, the communication system or any node or station for embodying the technique may further include any feature disclosed in the context of the first and / or second method aspects, and vice versa. Particularly, any one of the units and modules disclosed herein may be configured to perform or initiate one or more of the steps of the first and / or second method aspects.

[0134] Brief Description of the Drawings

[0135] Further details of embodiments of the technique are described with reference to the enclosed drawings, wherein:

[0136] Fig. 1 shows a schematic block diagram of an embodiment of a device for operating a wireless access network, particularly for mitigating an adversarial attack on a Distributed Multiple-Input Multiple-Output (D- MIMO) network;

[0137] Fig. 2 shows a schematic block diagram of an embodiment of a device for deploying a wireless access network, particularly for a D-MIMO network capable of mitigating an adversarial attack;

[0138] Fig. 3 shows a flowchart for an embodiment of a method of operating a wireless access network, which method may be implementable by the device of Fig. 1;

[0139] Fig. 4 shows a flowchart for an embodiment of a method of deploying a wireless access network, which method may be implementable by the device of Fig. 2;

[0140] Fig. 5 schematically illustrates a first embodiment of a radio access network as an example of a wireless access network comprising an embodiment of the device of Fig. 1 for performing the method of Fig. 3 for determining an AP;

[0141] Fig. 6 schematically illustrates a radio access network comprising an algorithmic CP for determining a beam; Fig. 7 schematically illustrates a second embodiment of a radio access network as an example of a wireless access network comprising an embodiment of the device of Fig. 1 for performing the method of Fig. 3 for determining a beam;

[0142] Fig. 8 schematically illustrates a conventional CP including an Al model (i.e., a ML system) without scrambler;

[0143] Fig. 9 schematically illustrates an attack on the Al model of the conventional CP;

[0144] Fig. 10 schematically illustrates an architecture, which is asymmetric with respect to permuting the APs, for an embodiment of the Al model of the CP;

[0145] Fig. 11 schematically illustrates a reduced architecture, which is also asymmetric with respect to permuting the APs, for a surrogate Al model of an attacker;

[0146] Fig. 12A schematically the detrimental effect of an attack in terms of RSRP loss as a function of the percentile of UEs;

[0147] Fig. 12B schematically the detrimental effect of an attack in terms of RSRP loss as a function of perturbation power level;

[0148] Fig. 13 schematically illustrates an embodiment of the CP including an Al model (i.e., a ML system) with scrambler;

[0149] Fig. 14 schematically illustrates a first embodiment of a communication system comprising an embodiment of the device of Fig. 1 for performing the method of Fig. 3 and an embodiment of the device of Fig. 2 for performing the method of Fig. 4;

[0150] Fig. 15 schematically illustrates a second embodiment of a communication system comprising an embodiment of the device of Fig. 1 for performing the method of Fig. 3 and an embodiment of the device of Fig. 2 for performing the method of Fig. 4;

[0151] Fig. 16A shows a flow chart for a first embodiment of the method of Fig. 4;

[0152] Fig. 16B shows a flow chart for a second embodiment of the method of Fig. 4;

[0153] Fig. 17 shows a schematic block diagram of a relay radio device embodying the device of Fig. 1;

[0154] Fig. 18 shows a schematic block diagram of a radio access network embodying the device of Fig. 2;

[0155] Fig. 19 schematically illustrates an example telecommunication network performing an embodiment of the methods of Figs. 3 and 4, optionally connected via an intermediate network to a host computer;

[0156] Fig. 20 shows a generalized block diagram of a host computer communicating via a base station or radio device functioning as a gateway with a user equipment over a partially wireless connection; and

[0157] Figs. 21 and 22 show flowcharts for methods implemented in a communication system including a host computer, a base station or radio device functioning as a gateway and a user equipment.

[0158] Detailed Description

[0159] In the following description, for purposes of explanation and not limitation, specific details are set forth, such as a specific network environment in order to provide a thorough understanding of the technique disclosed herein. It will be apparent to one skilled in the art that the technique may be practiced in other embodiments that depart from these specific details. Moreover, while the following embodiments are primarily described for a New Radio (NR) or 5G implementation, it is readily apparent that the technique described herein may also be implemented for any other radio communication technique, including a Wireless Local Area Network (WLAN) implementation according to the standard family IEEE 802.11, 3GPP LTE (e.g., LTE-Advanced or a related radio access technique such as MulteFire), for Bluetooth according to the Bluetooth Special Interest Group (SIG), particularly Bluetooth Low Energy, Bluetooth Mesh Networking and Bluetooth broadcasting, for Z-Wave according to the Z-Wave Alliance or for ZigBee based on IEEE 802.15.4.

[0160] Moreover, those skilled in the art will appreciate that the functions, steps, units and modules explained herein may be implemented using software functioning in conjunction with a programmed microprocessor, an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), a Digital Signal Processor (DSP) or a general purpose computer, e.g., including an Advanced RISC Machine (ARM). It will also be appreciated that, while the following embodiments are primarily described in context with methods and devices, the invention may also be embodied in a computer program product as well as in a system comprising at least one computer processor and memory coupled to the at least one processor, wherein the memory is encoded with one or more programs that may perform the functions and steps or implement the units and modules disclosed herein.

[0161] Fig. 1 schematically illustrates a block diagram of an embodiment of a device for operating a wireless access network according to the first device aspect. The device is generically referred to by reference sign 100.

[0162] The device 100 comprises a measurement module 102 configured to obtain measurement values indicative of wireless links between a wireless device (WD) and multiple APs distributed in the wireless access network.

[0163] The device 100 further comprises a service module 104 configured to serve or initiate serving the WD using at least one AP among the APs. The at least one AP is determined using a central processor (CP) comprising a scrambler and a machine learning system (ML system). The scrambler applies the obtained measurement values to an input layer of the ML system according to a permutation of the multiple APs, and an architecture of the ML system is asymmetric with respect to the permutation of the multiple APs.

[0164] Any of the modules of the device 100 may be implemented by units configured to provide the corresponding functionality. The device 100 may also be referred to as, or may be embodied by, the CP. The CP 100 and the APs of the wireless access network and / or the node of the core network may be in direct (e.g., radio) communication, e.g., at least for the obtaining of the measurements and / or the sending of the configuration message. The core node may be embodied by the below device 200.

[0165] Fig. 2 schematically illustrates a block diagram of an embodiment of a device for controlling a wireless access network according to the second device aspect. The device is generically referred to by reference sign 200.

[0166] The device 200 comprises a determination module 202 that determines a permutation of multiple APs distributed in the wireless access network. Herein, the permutation is applied prior to a machine learning system (ML system) that processes measurement values indicative of wireless links between a wireless device (WD), and the multiple APs. An architecture of the ML system is asymmetric with respect to the permutation of the multiple APs. The device 200 further comprises a configuration module 204 that sends a configuration message indicative of the determined permutation to a central processor (CP) configured to determine at least one AP, wherein a scrambler of the CP applies the measurement values to an input layer of the ML system of the CP according to the permutation.

[0167] Any of the modules of the device 200 may be implemented by units configured to provide the corresponding functionality.

[0168] The device 200 may also be referred to as, or may be embodied by, the core node (or briefly: core network). The core node 200 and the CP may be in (e.g., direct radio communication), e.g., at least for the obtaining of the measurements and / or the sending of the configuration message. The core node (CP) may be embodied by the above device 100.

[0169] Fig. 3 shows an example flowchart for a method 300, e.g. of operating a wireless access network. The method may embody the first method aspect and / or may comprise the steps indicated in Fig. 3.

[0170] The method 300 may be performed by the device 100. For example, the modules 102 and 104 may perform the steps 302 and 304, respectively. Fig. 4 shows an example flowchart for a method 400, e.g. of controlling a wireless access network. The method may embody the second method aspect and / or may comprise the steps indicated in Fig. 4.

[0171] The method 400 may be performed by the device 200. For example, the modules 202 and 204 may perform the steps 402 and 404, respectively.

[0172] In any aspect, the technique may be applied to uplink (UL), downlink (DL) or direct communications between radio devices, e.g., device-to-device (D2D) communications or sidelink (SL) communications.

[0173] Each of the APs or CP 100 or node 100 of the wireless access network may be a base station. The core node 200 of the core network may be a functional unit of the core network or a cloud implementation.

[0174] For brevity, and without limitation thereto, the WD is described herein as a radio device or UE, and the wireless access network is described as a radio access network (RAN), e.g., according to 3GPP 4G LTE or 5G NR.

[0175] Herein, any radio device may be a mobile or portable station and / or any radio device wirelessly connectable to a base station or RAN, or to another radio device. For example, the radio device may be a user equipment (UE), a device for machine-type communication (MTC) or a device for (e.g., narrowband) Internet of Things (loT). Two or more radio devices may be configured to wirelessly connect to each other, e.g., in an ad hoc radio network or via a 3GPP SL connection. Furthermore, any base station may be a station providing radio access, may be part of a radio access network (RAN) and / or may be a node connected to the RAN for controlling the radio access. For example, the base station may be an access point, for example a Wi-Fi access point.

[0176] Herein, whenever referring to noise or a signal-to-noise ratio (SNR), a corresponding step, feature or effect is also disclosed for noise and / or interference or a signal-to-interference-and-noise ratio (SINR).

[0177] In any embodiment, the scrambling scheme or permutation may be predefined, e.g. may be stored (e.g., in the CP or the AP) in memory, or hard-coded or hard- wired in the transmitting wireless device, or preconfigured or configured by a node of the wireless access network or a node of the core network.

[0178] Herein, a list of the form A, B, and / or C (also written as A, B and / or C) may correspond to at least one or each of A, B, and C, i.e., A and / or B and / or C. Any aspect of the technique may be implemented as a method to mitigate adversarial attacks in an Al-driven (i.e., ML-driven) wireless access network, e.g. for Al-driven beam selection in a D-MIMO network.

[0179] The ML system (i.e., the Al model) of the CP may comprise one or more Deep Neural Networks (DNNs). Alternatively or in addition, embodiments may use or avoid Adversarial Machine Learning for the ML system.

[0180] In any embodiment, the determining by using the ML system may be a beam selection.

[0181] At least some embodiment may improve the security of the wireless access network. In a variant of any embodiment, the wireless access network may be a D-MIMO RAN and / or a sixth generation (6G) access network.

[0182] The distributed multiple-input multiple-output (D-MIMO) network may be a type of wireless access network for communication systems beyond fifth generation (B5G), wherein many radio units (RUs, as examples of the APs) are geographically distributed in a region (as an example of the D-MIMO site) to increase coverage and / or reliability. The results obtained in previous research studies show the benefits of D-MIMO compared to traditional uncoordinated small-cells, which are used in all previous generations, and collocated massive MIMO which is currently used in 5G systems. In indoor mobile network deployments for factories, warehouses, and offices focusing on applications with high demands of network performance like extended reality (XR) and dense machine-type communication (MTC), D-MIMO can enhance capacity, reliability, and resilience. D-MIMO is well- suited for high-frequency bands with challenging radio propagation, thanks to its high line-of-sight probability and increased macro diversity.

[0183] Macro diversity is a technique used in wireless communication systems to improve the quality and reliability of the signal by combining multiple copies of the same signal received at different base stations. In the context of Distributed Multiple Input Multiple Output (D-MIMO), macro diversity can be achieved through the use of multiple base stations, each equipped with multiple antennas, to transmit and receive signals. This allows for the creation of multiple communication paths between the transmitter and receiver, increasing the reliability and capacity of the system. The use of macro diversity in D-MIMO systems can also help mitigate the effects of fading and interference, improving the overall performance of the system.

[0184] Fig. 5 schematically illustrates an example of a D-MIMO network 500 with 16 RUs 502 which are all connected to an embodiment of the CP 100 (also referred to as Central Processor Unit or CPU), e.g. at the edge of the wireless access network. The CP 100 includes the ML system 103, which is also referred to as Al model or Al agent or briefly Al or ML.

[0185] Preferably, the ML system 103 comprises at least one neural network which architecture is not symmetric with respect to permuting the multiple Al input features of the ML system 103.

[0186] There are lots of functions that are needed to be performed in order to orchestrate the network resources (e.g., determine at least one to radio resource management, RRM, parameter), e.g. to optimize the network and / or maximize spectral efficiency of the served WDs 504 (also referred to as UEs or users).

[0187] In D-MIMO, at least one or most or all of these functions are performed (e.g. corresponding parameters are determined) by the ML system due to various reasons such as reducing complexity and overhead. Among these functions, power control, beam sweeping / selection and RU selection are some of the most important Al-driven tasks for D-MIMO.

[0188] To explain them shortly, the aim of power control is to allow RUs allocate their power among users to optimize the system's performance. There are important benefits for applying power control techniques in D-MIMO such as interference management, energy efficiency and connectivity management. And, although there are many RUs 502 in a D-MIMO network, using all RUs to serve a single UE might not be practical due to high fronthaul load. A better approach is to find a suitable subset of RUs to serve each UE. And this is where the Al techniques come into play. Embodiments may use the ML system 103 to efficiently find user- centric RU sets (i.e., determine the at least one AP 502) to serve UEs 504 according to the step 304. And finally, as an alternative to complete beam sweeping, a subset of all possible beams can be swept, and the one or more best beams can be determined by using the ML system 103, e.g. to decrease the beam training overhead. This functionality is called an Al-driven beam sweeping / selection task. In summary, there is a need for Al functionalities at the edge (i.e. at the CP or central processing unit, CPU) that is taking care of these important tasks. Herein, the terms central processing unit (CPU) and central processor (CP) are used interchangeably. The term CPU is not related to or not limited to the similar expression for a microprocessor.

[0189] In any embodiment the serving step 304 may comprise determining at least one beam, i.e., a beam selection (also referred to as a task or process), for D-MIMO.

[0190] Beam selection for D-MIMO is a process to predict the best beam and access point (AP) to serve the user equipment (UE) in a D-MIMO network using a subset of beam measurements in downlink (DL). In this document, the terms access point (AP) and radio unit (RU) are used interchangeably. It can be applied to find the initial wide beam (using SSBs) or find the narrow data beam (using CSI-RSs). In this invention, we focus on initial access where wide beams are sent by the network including SSBs to transmit common information to UEs to make the initial access possible. As SSBs include common information for synchronization and system information, there is no UE specific beam at this stage. Each beam can be received and decoded by any UE around the APs in a parallel way. Therefore, we only consider a single UE throughout the document.

[0191] Conventionally finding the best AP 504 and beam requires measuring the DL channel for all the APs using all the beams as in Fig. 6. In D-MIMO, this becomes resource-heavy, especially at millimeter Wave (mmWave).

[0192] Fig. 6 schematically illustrates a conventional (non-AI-driven) beam sweeping process for D-MIMO. In this example, each AP has 4 different beams to transmit. In the standard approach, all possible 6 x 4 = 24 beams are sequentially transmitted to UEs to determine the best beam.

[0193] To overcome the above problem, Al-based determining is used in the step 304. The aim of Al-driven beam selection is to investigate whether best AP 502 and / or beam 130 can be inferred from sounding the channel (i.e., the obtained measurement values) on a subset of APs (i.e., the multiple APs 504) and beams (e.g., a subset of the combination of APs and beams, optionally including all APs of the D-MIMO site), with the help of the ML system 103, e.g. as shown in Fig. 7.

[0194] Fig. 7 schematically illustrates Al-driven beam selection as an example of the determining using the ML system according to the step 304. In this example, only 7 beams out of 24 are transmitted by the network 500. The RSRP measurements (as examples of the measurement values) obtained in the step 302 for these 7 beams are input to an Al model to predict the best beam 130 (i.e., the determined beam 130). The predicted beam 130 can be among the ones which are not sent (i.e., transmitted).

[0195] In D-MIMO, link quality in each beam is measured by the UE 504 on downlink reference signals (DL RSs). And then, the UE 504 provides a link quality report in the uplink (UL) according to the step 302. So, the input of the Al model 103 is the measured RSRP values (i.e., the measurement values). And the output of the Al model can be the index of the best beam (e.g., as a classification task). The basic functionality of the Al model 103 is shown in Fig. 8. This may correspond to a conventional CP 10 without a scrambler.

[0196] Fig. 8 schematically illustrates an example I / O of the Al model (i.e., the ML system) 103 for the determining, here for beam selection as an example. In this example, there are 4 x 4 = 16 beams in total, but the network 500 sends a proper subset 8 of them. The trained Al model 103 uses the measured RSRP values for these 8 beams to predict the index of the best beam 130. As can be seen, the predicted beam 130 is among the ones which are not sent.

[0197] The conventional CP without scrambler uses an Al model (i.e., ML system) 130, which have recently been discovered to be vulnerable against malicious attacks. In fact, very small and often undetectable changes (i.e., permutations) in data samples (e.g., the measurement values) are enough to fool state-of-the-art classifiers in inference time and lead to incorrect predictions 130. The past few years have witnessed extensive research, which shows the vulnerability of Al- driven systems 103 in different domains. And despite the distributed nature of the communication domain and the heterogeneity of the network 500, there is still a risk of adversarial attacks in a telco environment. Previously mentioned research on "Practical Adversarial Attacks Against Al-Driven Power Allocation in a Distributed MIMO Network" indicates that adversarial attacks with optimized perturbations can target Al-driven tasks such as power control and degrade the performance of a D-MIMO network in terms of both spectral and energy efficiency.

[0198] For concreteness and not limitation, this detailed description focuses on Al- driven beam selection task in D-MIMO, and the inventors have shown that a malicious software in a compromised UE 504 can easily target the beam selection model 103 in a conventional CPU 10 and result in dramatic performance decreases.

[0199] As reported in news article "Qualcomm vulnerability impacts nearly 40% of all mobile phones" (available on bleepingcomputer.com, published May 6, 2021), some UE modems are vulnerable to malicious software. A highly capable malware can settle down in a UE and read and / or modify RSRP values measured by the UE modem. The malware can train a local Al model 103' that imitates the actual Al model 103 used in convention CPU 10 to carefully craft adversarial perturbations. In Fig. 9, a typical attack scenario is shown.

[0200] Fig. 9 schematically illustrates a block diagram of a typical adversarial attack scenario.

[0201] Beam selection for D-MIMO is a process to predict the best beam and access point (AP) to serve the user equipment (UE) 504 in a D-MIMO network 500 using a subset of beam measurements in downlink (DL). In initial access, wide beams are sent by the network including Synchronization Signal Blocks (SSBs) to transmit common information to UEs to make the initial access possible. As SSBs include common information for synchronization and system information, there is no UE specific beam at this stage. Each beam can be received and decoded by any UE around the APs in a parallel way. Each UE 504 in the coverage area can measure the link quality in each beam on DL RS. So, the UEs will know the corresponding RSRP values for each beam (UEs are providing a link quality report to the network in the UL and the network can have a chance to train its own model using this info). And also, since the data transmission will be enabled by the network using the best beam, the UEs 504 will also have a chance to learn the best beam information. Considering the scenario where a compromised UE 504 spends enough time in the network coverage area by dynamically moving around, the UE will have a chance to collect enough information to train a surrogate Al model 103'.

[0202] Here, for illustration and not limitation, i.e., to make the attack practical, we assume the following:

[0203] 1) The attacker has no access to actual Al model in CPU 10 (i.e., no so-called white-box attack). So, it trains a separate surrogate model 103' imitating the functionality of the actual Al model 103 and crafts the perturbations in a blackbox manner.

[0204] 2) The attacker designs a perturbation so that for each measurement value (e.g., RSRP value), the added perturbation has a magnitude of at most e dB where e is a small quantity. This is required for attacker not to be detected by the network. For example, it should be smaller than typical large-scale channel variations, such as shadowing standard deviation.

[0205] 3) The RSRP reports are sent in UL according to the step 302 so that in each report only the RSRP values corresponding to the beams of the last AP are included. This is compatible with what is in the current 5G NR implementation. Due to this assumption, the attacker crafts the adversarial samples with missing information (to craft adversarial samples for the m-th AP, only the RSRP values for 1, 2, ..., m-th AP's beams can be used.) As the best perturbation calculation requires all RSRP information for all APs one at a time, we assume that the attacker uses a technique called universal adversarial perturbation (UAP) allowing to craft perturbations with limited knowledge.

[0206] 4) Due to UL channel effects, the measured RSRP values might not be perfectly delivered to the network. We add some noise to the RSRP values (in both clean and perturbed cases) to model this effect.

[0207] As shown in Table I, a typical adversarial attack can cause the Al model in CPU to select a beam whose RSRP value is much less than the actual best beam. In this table "Clean" shows the Al performance without any attack, "WGN" shows the simple white Gaussian noise attack that randomly generates the perturbation values without any calculation, and UAP shows the universal adversarial perturbation method that crafts the perturbation values using the gradient of the loss function of the surrogate Al model.

[0208] Table I: The mean RSRP differences w.r.t. the best beam for various cases. In this experiment, there are 10 RUs each with 32 beams in total. The network sends only 80 out of 320 beams to decrease the overhead by 75%.

[0209] As shown by Table I, in the clean case where there is no attack, Al performs very well. It can find a beam whose RSRP value is only 0.08 dB less than the best beam. Although WGN attack degrades the performance, its effect is limited. On the other hand, under adversarial attack (UAP) in black box setting, there is a significant performance loss. Attacker can degrade the communication signal strength more than 6 dB by adding a small enough perturbation. In general, for outdoor case, shadowing standard deviation is assumed to be 8 dB and any perturbation less than that value has a small probability to be detected.

[0210] In numerical experiments, the inventors used below model architectures for the original Al model 103 in CP 100 and surrogate Al model 103' in compromised UEs 504, and both models achieved considerable performance levels. So, we can safely assume that training a surrogate Al model 103' will be a very straightforward task for a malicious actor.

[0211] Fig. 10 schematically illustrates an embodiment of the Al model 103 in the CP 100. At least the input layer 101 is asymmetric under permutations of the APs 502 (i.e., permuting the order of the AP input features, e.g. wherein each AP input feature comprises the measurement values belonging to the same AP). In this example, the input layer 101 comprises a 2D convolutional neural network or Conv2D, which may comprise a ML module that performs a two-dimensional convolution operation on an input signal composed of several input planes. A convolution is a mathematical operation that combines two functions (e.g., the measurement values and the kernel) to produce a third function (i.e., the output of the input layer to the second layer), e.g. that expresses how the shape of one is modified by the other. In ML, a convolution is often used to extract features from images or other types of data.

[0212] The Conv2D module (i.e., the input layer) may comprise at least one of the following parameters that control how the convolution is performed, such as:

[0213] • Filters: the number of output channels (or feature maps) produced by the convolution.

[0214] • Kernel size: the height and width of the convolution window (or kernel) that slides over the input signal.

[0215] • Strides: the number of pixels that the convolution window moves along the height and width dimensions of the input signal.

[0216] • Padding: the amount of zero-padding added to the input signal before performing the convolution. This can be either "valid" (no padding), "same" (padding to keep the output size equal to the input size), or an integer or a tuple of integers specifying the padding on each side.

[0217] • Dilation rate: the spacing between the kernel points. This can be used to create dilated convolutions, which can capture larger receptive fields without increasing the kernel size.

[0218] • Groups: the number of groups to split the input and output channels into. This can be used to create depth-wise convolutions, which apply a single filter per input channel, or group convolutions, which reduce the computational cost by sharing filters across groups.

[0219] • Activation: an optional activation function to apply to the output of the convolution, such as Rectified Linear Unit (ReLU), sigmoid, or tanh.

[0220] • Use bias: a Boolean flag indicating whether to add a bias term to the output of the convolution.

[0221] • Kernel initializer, bias initializer, kernel regularize^ bias regularize^ activity regularize^ kernel constraint, bias constraint: optional initializers, regularizers, and constraints for the kernel and bias parameters. Fig. 11 schematically illustrates an architecture of the surrogate model 103', which is also asymmetric with respect to permuting the multiple APs 502.

[0222] We present the results under UAP attack that uses the surrogate model 103' given above and RSRP measurements done by the UE 504 for each RU 502 sequentially. Above Table I shows the effectiveness of the attack on the beam selection task. Furthermore, Figs. 12A and 12B show diagrams resulting from simulations, which show the effect of an adversarial attack on a wireless access network 500 using a conventional CP 10 without scrambler.

[0223] As can be seen from the Fig. 12A and Fig. 12B, adversarial attacks (shown as UAP) lead to a considerable degree of RSRP error for beam selection task. And the amount of error introduced by the adversarial attack is much larger than conventional attacks (applying white Gaussian noise, shown as WGN). The first figure 12A shows the RSRP error (RSRP difference with respect to best beam) statistics at e = 6 dB in black-box scenario using a surrogate model 103'. In second figure 12B, we see 90th percentile RSRP errors as a function various epsilon value.

[0224] The results indicate that the problem is indeed important, and a defense mechanism is required to mitigate such adversarial attacks.

[0225] Regarding the adversarial defense solutions, it is noted that there are different defense solutions proposed in literature for adversarial attacks. Most popular ones are adversarial training, defensive distillation, input denoising etc. However, there is still no existing adversarial defense mechanism that achieves both efficiency and effectiveness against adversarial samples. The most effective defense mechanism, which is adversarial training, is too computationally expensive for practical deployment, while many efficient heuristic defenses have been demonstrated to be vulnerable to adaptive adversaries. It is known that adversarially trained models can be made robust to whitebox attacks (i.e., with knowledge of the model parameters) if the perturbations computed during training closely maximize the model's loss. However, it is shown that adversarially trained models are still vulnerable to black-box adversarial attacks. Most of the practical attack scenarios in telecommunications or wireless domain will be blackbox. As a solution to mitigate blackbox adversarial attacks, Ensemble Adversarial Training is proposed. This method augments a model's training data with adversarial examples crafted on other static pre-trained models. By this way, it decouples adversarial example generation from the parameters of the trained model and increases the diversity of perturbations seen during training. However, ensemble adversarial training is an extremely expensive process, as one needs to train tens of different models using adversarial training and still not provide a perfect performance. There is also the defensive distillation method which is also a very famous one. And it is known that this solution does not work for blackbox attacks as well (and vulnerable to some whitebox attacks).

[0226] Any aspect of the subject technique may comprise, alone or in combination with embodiment disclosed herein, a method of mitigating adversarial effectiveness of an attacker using a scrambler 110. This defensive solution can significantly reduce the attack success without degrading clean (i.e., natural) performance.

[0227] Any embodiment of the technique may be based on selecting fixed indices for each AP 502 and shuffling the order of APs 502 (i.e., applying the scrambler 110) during both training of the ML system 103 (i.e., Al model training phase) and inference phase (i.e., the determining in the serving step 304) via the scrambler 110 prior to the Al model 103.

[0228] Preferably, embodiments use different scrambling patterns for each D-MIMO network deployments (e.g., different D-MIMO sites). In a variant of any embodiment, the selection of fixed indices can be same for different D-MIMO deployments. During the numerical experiments, different Al models have been trained. The best performance is achieved when selecting fix beam indices for each AP 502. So, choosing a fix set of beams for each AP has an additional benefit in terms of model performance.

[0229] One of the key points is to shuffle according to the permutation the ordering of APs in the input layer 101 of the Al model 103 so that no outsider can imitate the target Al model 103 (running in CPU 100) by training a surrogate model 103'. The other key point is to ensure the confidentiality of each chosen scrambler via different encryption techniques.

[0230] In a simplified symbolic consideration, the task of the attacker is to approximate a function for the determining (e.g., of the best beam 130). The task of the CP 100 including the scrambler 110 is to approximate another function F = H ° Sinv, wherein Sinv = Stranspose is the inverse of the permutation of the scrambler, so that CP 100 approximates the function H.

[0231] This task F of the network 500 (or CP 100) is hidden from the attacker that does not know S or Sinv. While the attacker does not have to solve the task F of the network but aims at solving H, the asymmetry of the model architecture of the ML system 103 (which action is denoted symbolically by MCP) with respect to the AP-permutations limits the solution space of CP in a different shape as compared to the surrogate model 103', so that even though

[0232] CP=MCp ° S = F ° S = H ° Sinv ° S = H, approximates the task of beam selection (which is also the task of the attacker), the sensitive points where the result of the determination switches are different for CP 100 and the surrogate Al model 103'. Thus, the perturbations crafted based on the surrogate Al model 103' fail to trigger a switch in the determination result of the CP 100 using the scrambler.

[0233] Preferably, in a variant of any embodiment, to be successful against adversarial attacks, the scrambling operation, i.e. the permutation S, is protected and not known or extracted by the attacker. To determine and protect the scrambling operation information to be used prior to the Al model 103 to mitigate the adversarial attacks, the following alternative and combinable approaches are taught:

[0234] Approach 1 (Non-data aided):

[0235] 1. Determination of a high number of different robust scramblers.

[0236] 2. A technique to map each scrambler to a different D-MIMO site.

[0237] 3. Protection of each scrambler used in D-MIMO sites (CPU) by encryption techniques.

[0238] Approach 2 (Data aided):

[0239] 1. We define a function to be placed in the core network or in the CPU of the related D-MIMO site that determines a robust scrambler.

[0240] 2. The function uses RSRP data collected from the field to find the optimal scrambler. Any embodiment of the method 300 may comprise at least one of the following steps in a method to mitigate effects of adversarial attacks (e.g., on beam selection and / or in initial access) in a D-MIMO network 500: a. A scrambling operation (i.e., a permutation) is applied in the step 304 to the AP / beam indices (e.g. fixed) before the Al operation (i.e., at the input layer of the ML system 103). b. A deep neural network-based Al model 103 is used in the step 304 to determine (e.g., estimate) the best AP / beam index using a set of measurement values (e.g., RSRP measurement reports) sent by UEs 504 in the UL. c. The technique is performed in an Al-based initial access scheme in D-MIMO, wherein a subset of all possible beams from all possible APs are sequentially sent and UEs report the measured RSRP values back to the network sequentially where each report includes RSRP values for a single AP.

[0241] Any embodiment of the method 400 may comprise at least one of the following steps in a method to determine and / or protect the scrambling operation information: a. A step 402 of determining a high number of robust scrambling operations (e.g., permutations). b. Using (e.g., by means of the configuration message) a different scrambler (i.e., different scrambler schemes) for each D-MIMO site. c. Sending the configuration message using encryption to increase the security of scrambler information used in a specific D-MIMO site.

[0242] In a variant of any embodiment of the method 400, the scrambler (e.g., the permutation) is determined specific to a D-MIMO site according to the positions of APs, e.g. using data-aided process, optionally including at least one of: a. A function placed in core network 200 or in the CPU 100 in the related D- MIMO site, which function determines optimal scrambler according to the RSRP data collected from the field. b. A training process enabling to optimize the scrambler in the specific D- MIMO site.

[0243] Embodiments of the technique can be regarded as a method of hiding the original Al model 103 from outside. Since the original model 103 (including its sensitive points) will not be known by anyone via the operation of the scrambler 110, it will not be possible for someone to apply white-box attacks or train a surrogate model 103' to craft perturbation in black-box manner. In other words, embodiments of the technique makes it impossible for the attacker to perform adversarial attacks and eliminates the need to employ previously mentioned adversarial defense methods.

[0244] It is remarked that a similar operation would not be feasible in other domains like image, text or audio as the scrambling of the input will completely change the input meaning and break the correlation between neighbor pixels, whereas the subject technique can be useful for Al-driven wireless access networks. While the presented numerical experiments performed beam selection task as the determining in the serving step 304, other embodiments are also valid for other tasks like Al-driven power control task.

[0245] Thus, at least some embodiments of the defense technique have at least one of the following advantages. Firstly, the proposed methods have a very low computational complexity. It only includes a scrambling operation (i.e., the permutation, which is well-protected) to shuffle (i.e., permute) the AP indices before running the Al model 103 in the CP 100. Secondly, it significantly enhances the beam selection performance under adversarial attacks. As shown in below Table II, it can enhance the signal strength more than 4 dB under a UAP attack with 6 dB s value. Thirdly, it does not affect the clean performance. As the beam selection problem is symmetric w.r.t. APs, by introducing a scrambling before the Al will not affect the beam selection performance.

[0246] Table II: The mean RSRP differences w.r.t. best beam for under attack and by the proposed defense method. Above Table II presents the performance results from numerical experiments for WGN, UAP (without any defense) and the proposed defense technique (scrambling prior to Al, e.g. wherein a scrambler with the new AP order 6, 10, 5, 1, 4, 9, 3, 8, 2, 7 is used) under various e values. In this experiment, there are 10 APs each with 32 beams. The system selects 80 fixed beams (8 beams from each AP) to transmit the SSBs and get the RSRP information. We observe that the defense technique can increase the performance by more than 4 dB for relatively small e values. The effect of the adversarial attack when the defense method is available becomes less than the standard WGN (when there is no defense) showing the effectiveness of our proposed technique.

[0247] Regarding the difficulty to train a surrogate Al model 103' for the attacker if we have the scrambler 110 in place, it would be almost impossible (as long as the scrambling information (i.e., the permutation) is kept secret and unknown by anyone apart from the CP 100). Assuming that the number of beams used for selection task is M, then there would be Ml different scrambling options for the CP 100 and the probability of choosing the correct scrambler will be 1 / M for the adversary. So, for example assuming M > 20, the probability will be extremely low (almost zero) and for the attacker to be successful, the adversary needs to train as much as Ml different surrogate Al models 103', which is also practically impossible.

[0248] While a system comprising devices 100 and 200 is described in more detail hereinbelow, the skilled person understands the features of such a system are disclosed for a pertinent one and both of the devices 100 and 200.

[0249] A scrambler 110 is used prior to the Al model 103 to shuffle (i.e., permute) the places of APs 502 (e.g., at the input layer 101) so that the attacker cannot design an effective perturbation to fool the Al model 103. A reason behind this solution can be explained as follows.

[0250] The Al solution space includes some sensitive points (in the direction of the gradient of the objective loss function) where the output rapidly changes with respect to the changes in input. The adversarial attack aims at designing a small perturbation to push the input towards those sensitive points (by aiming to maximize the objective loss function in gradient direction). By applying the scrambler 110 (and retraining the Al model 103 accordingly), the locations of those sensitive points will change, making the attack inefficient (assuming that the adversary does not have the knowledge of this scrambling operation.)

[0251] Fig. 13 present a block diagram of an embodiment of the CP 100 including the scrambler 110. Fig. 13 shows an example scrambling scheme (i.e., the permutation) and other type of scramblers can also be used.

[0252] In the embodiment of the CP 100 illustrated in Fig. 13, beam indices 2 and 4 are selected as fixed indices for each AP 502, but the ordering of the APs 502 has been randomly shuffled. This operation has been followed for both the model training and inference phases.

[0253] It is required for this method to enhance security of beam selection that the scrambler information should be protected against the attackers. Optionally, to determine a robust scrambler and increase the security of scrambler information the following two approaches are taught, which may be implemented alone or in combination.

[0254] Approach 1: (Non-data aided)

[0255] Fig. 16A shows a flowchart of an implementation of the method 400.

[0256] 1. Determination of a high number of different robust scramblers:

[0257] In current 5G NR standard different cells are identified by a unique cell ID information determined by PSS and SSS indices used in SSB block. There are 3 different PSS and 336 different SSS sequences indicating 1008 different cell IDs. Assuming a similar identification in D-MIMO, we can aim to create 1008 different scramblers to be used in corresponding D-MIMO sites. To find such scramblers, we propose the following operation.

[0258] Repeat the following steps until finding 1008 different scramblers: a. Randomly generate a scrambler. b. Make a robustness analysis by generating adversarial samples and measuring the selected beam RSRP enhancements obtained by using the generated scrambler. c. If the average enhancement is larger than a predetermined threshold, then save the scrambler. 2. A technique to map each scrambler to a different D-MIMO site:

[0259] To map those scramblers to each D-MIMO site, we propose to define a one-to- one mapping which is a separate function, encrypted and located in the core network.

[0260] 3. Protection of each scrambler used in D-MIMO sites by encryption techniques:

[0261] To increase the security of scrambler information, we propose not to place a one- to-one mapping (e.g., as described in above section) into each specific D-MIMO site. Instead, we propose to use known encryption techniques over backhaul links, i.e., the connection between CPU and core network. The core network can transmit the scrambler ID over backhaul links to the specific D-MIMO site via encryption when needed. According to 3GPP standardizations specified in the 3GPP documents TS 33.310 and TS 33.320, the establishment of secure backhaul link is based on IKEv2. Thus, backhaul links provide integrity, confidentiality and replay protection of the transmitted data. This makes for UE or fronthaul originated attacker to get this scrambler ID information impossible.

[0262] Fig. 14 schematically illustrates a block diagram of for protecting the scrambler information (i.e., the scrambling scheme, particularly the permutation). Here, a scrambler codebook (e.g., comprising many different robust scramblers) is used in both the core network 200 and the CPU 100. The scrambler ID represents a one-to-one mapping between the ID (e.g., included in the configuration message) and a specific scrambler (i.e., a specific scrambling scheme). This way, the confidentiality of the scrambler can be increased and signaling overhead can be reduced.

[0263] The ID may be uniquely indicative of the permutation, e.g., the mapping between labels of the AP input features at the input layer 101 and the APs 502.

[0264] Approach 2: (Data aided)

[0265] A flowchart summarizing the steps for an implementation of the method 400 is shown in Fig. 16B. In this approach, we first define a function in the core network or in the CPU 100 of the related D-MIMO site that will determine a robust scrambler. This function is fed by collected RSRP data from the field. This data includes some correlation according to the network deployment, i.e., positions of APs. This approach can theoretically find the best scrambler that fits best the spatial positions of the APs.

[0266] The following steps are taught to find a good scrambler (i.e., permutation):

[0267] 1. Start with a random scrambler (or the scrambler found by Approach 1).

[0268] 2. Define a cost metric using the RSRP difference between the best beam and the beam selected by the Al method (with the given scrambler) under adversarial attacks.

[0269] 3. Using any iterative classifier technique such as genetic algorithm, particle swarm optimization, etc., update the scrambler so that the cost metric enhances.

[0270] 4. Operate the algorithm until convergence.

[0271] This process provides a robust and site-specific scrambler, which is inherently protected against the attacker as it is an output of an algorithm.

[0272] An embodiment of the system 1400 includes the blocks as schematically illustrated in Fig. 15. The embodiments of the device 100 and the device 200 may comprise at least one of the following three functional blocks:

[0273] 1. A scrambler 110 (to make the Al model-based determination robust against attackers),

[0274] 2. A one-to-one mapper (to use different scrambler at each D-MIMO site), and

[0275] 3. A scrambler optimizer (to use the best scrambler at each D-MIMO site)

[0276] In one embodiment, for each D-MIMO site, there needs to be one site-specific Al model 103, e.g. for beam selection task (as an example of determining the RRM parameter value). That model needs to be trained by using site-specific data using the respective scrambler 110.

[0277] Preferably, one Al model architecture is used for each site. But for training and using this site-specific Al model, the scrambling operation is used. As a result, embodiments have at each site a single scrambler and a single Al model in operation. Any embodiment of the device 100 or 200 may be implemented at an access point (AP), central processing unit (CPU or CP), a radio unit (RU), a distributed unit (DU), or core network (CN).

[0278] Fig. 17 shows a schematic block diagram for an embodiment of the device 100. The device 100 comprises processing circuitry, e.g., one or more processors 1704 for performing the method 300 and memory 1706 coupled to the processors 1704.

[0279] For example, the memory 1706 may be encoded with instructions that implement at least one of the modules 102 and 104.

[0280] The one or more processors 1704 may be a combination of one or more of a microprocessor, controller, microcontroller, central processing unit, digital signal processor, application specific integrated circuit, field programmable gate array, or any other suitable computing device, resource, or combination of hardware, microcode and / or encoded logic operable to provide, either alone or in conjunction with other components of the device 100, such as the memory 1706, base station functionality or RAN functionality or CP functionality. For example, the one or more processors 1704 may execute instructions stored in the memory 1706. Such functionality may include providing various features and steps discussed herein, including any of the benefits disclosed herein. The expression "the device being operative to perform an action" may denote the device 100 being configured to perform the action.

[0281] As schematically illustrated in Fig. 17, the device 100 may be embodied by a node 1700 of a wireless access network (e.g., RAN), e.g., functioning as a base station or a CP. The transmitting station 1700 comprises a (e.g., backhaul) interface 1702 coupled to the device 100 for (e.g., radio) communication with one or more nodes of a core network, e.g., functioning as the device 200.

[0282] Fig. 18 shows a schematic block diagram for an embodiment of the device 200. The device 200 comprises processing circuitry, e.g., one or more processors 1804 for performing the method 400 and memory 1806 coupled to the processors 1804.

[0283] For example, the memory 1806 may be encoded with instructions that implement at least one of the modules 202 and 204. The one or more processors 1804 may be a combination of one or more of a microprocessor, controller, microcontroller, central processing unit, digital signal processor, application specific integrated circuit, field programmable gate array, or any other suitable computing device, resource, or combination of hardware, microcode and / or encoded logic operable to provide, either alone or in conjunction with other components of the device 200, such as the memory 1806, core network functionality. For example, the one or more processors 1804 may execute instructions stored in the memory 1806. Such functionality may include providing various features and steps discussed herein, including any of the benefits disclosed herein. The expression "the device being operative to perform an action" may denote the device 200 being configured to perform the action.

[0284] As schematically illustrated in Fig. 18, the device 200 may be embodied by a node 1800 of the core network, e.g., a session management function (SMF). The node 1800 comprises a (e.g., backhaul) interface 1802 coupled to the device 200 for (e.g., radio) communication with one or more nodes of the wireless access network (e.g., RAN), e.g., functioning as a CP.

[0285] With reference to Fig. 19, in accordance with an embodiment, a communication system 1900 includes a telecommunication network 1910, such as a 3GPP-type cellular network, which comprises an access network 1911, such as a radio access network, and a core network 1914. The access network 1911 comprises a plurality of base stations 1912a, 1912b, 1912c, such as NBs, eNBs, gNBs or other types of wireless access points, each defining a corresponding coverage area 1913a, 1913b, 1913c. Each base station 1912a, 1912b, 1912c is connectable to the core network 1914 over a wired or wireless connection 1915. A first user equipment (UE) 1991 located in coverage area 1913c is configured to wirelessly connect to, or be paged by, the corresponding base station 1912c. A second UE 1992 in coverage area 1913a is wirelessly connectable to the corresponding base station 1912a. While a plurality of UEs 1991, 1992 are illustrated in this example, the disclosed embodiments are equally applicable to a situation where a sole UE is in the coverage area or where a sole UE is connecting to the corresponding base station 1912.

[0286] Any of the base stations 1912 may embody the APs 502 and / or the CP 100. The access network 1911 may embody the wireless access network 500. The core network 1914 may embody the device 200. The telecommunication network 1910 is itself connected to a host computer 1930, which may be embodied in the hardware and / or software of a standalone server, a cloud-implemented server, a distributed server or as processing resources in a server farm. The host computer 1930 may be under the ownership or control of a service provider, or may be operated by the service provider or on behalf of the service provider. The connections 1921, 1922 between the telecommunication network 1910 and the host computer 1930 may extend directly from the core network 1914 to the host computer 1930 or may go via an optional intermediate network 1920. The intermediate network 1920 may be one of, or a combination of more than one of, a public, private or hosted network; the intermediate network 1920, if any, may be a backbone network or the Internet; in particular, the intermediate network 1920 may comprise two or more sub-networks (not shown).

[0287] Optionally, the host computer 1930 embodies the device 200.

[0288] The communication system 1900 of Fig. 19 as a whole enables connectivity between one of the connected UEs 1991, 1992 and the host computer 1930. The connectivity may be described as an over-the-top (OTT) connection 1950. The host computer 1930 and the connected UEs 1991, 1992 are configured to communicate data and / or signaling via the OTT connection 1950, using the access network 1911, the core network 1914, any intermediate network 1920 and possible further infrastructure (not shown) as intermediaries. The OTT connection 1950 may be transparent in the sense that the participating communication devices through which the OTT connection 1950 passes are unaware of routing of uplink and downlink communications. For example, a base station 1912 need not be informed about the past routing of an incoming downlink communication with data originating from a host computer 1930 to be forwarded (e.g., handed over) to a connected UE 1991. Similarly, the base station 1912 need not be aware of the future routing of an outgoing uplink communication originating from the UE 1991 towards the host computer 1930.

[0289] By virtue of the method 300 being performed by the core network 1914 and / or by virtue of the method 200 being performed by the wireless access network 1911, e.g., the base stations 1912 as APs 502 or CP, the performance or range of the OTT connection 1950 can be improved, e.g., in terms of increased throughput and / or reduced latency. More specifically, the host computer 1930 may indicate to the RAN 500, 1911, e.g. the base station 1912 (e.g., on an application layer) a Quality of Service (QoS) of the traffic or other parameter (e.g., a security or reliability level) that triggers the D-MIMO operation and / or the scrambling operation in the CP.

[0290] Example implementations, in accordance with an embodiment of the UE, base station and host computer discussed in the preceding paragraphs, will now be described with reference to Fig. 20. In a communication system 2000, a host computer 2010 comprises hardware 2015 including a communication interface 2016 configured to set up and maintain a wired or wireless connection with an interface of a different communication device of the communication system 2000. The host computer 2010 further comprises processing circuitry 2018, which may have storage and / or processing capabilities. In particular, the processing circuitry 2018 may comprise one or more programmable processors, application-specific integrated circuits, field programmable gate arrays or combinations of these (not shown) adapted to execute instructions. The host computer 2010 further comprises software 2011, which is stored in or accessible by the host computer 2010 and executable by the processing circuitry 2018. The software 2011 includes a host application 2012. The host application 2012 may be operable to provide a service to a remote user, such as a UE 2030 connecting via an OTT connection 2050 terminating at the UE 2030 and the host computer 2010. In providing the service to the remote user, the host application 2012 may provide user data, which is transmitted using the OTT connection 2050. The user data may depend on the location of the UE 2030. The user data may comprise auxiliary information or precision advertisements (also: ads) delivered to the UE 2030. The location may be reported by the UE 2030 to the host computer, e.g., using the OTT connection 2050, and / or by the base station 2020, e.g., using a connection 2060.

[0291] The communication system 2000 further includes a base station 2020 provided in a telecommunication system and comprising hardware 2025 enabling it to communicate with the host computer 2010 and with the UE 2030. The hardware 2025 may include a communication interface 2026 for setting up and maintaining a wired or wireless connection with an interface of a different communication device of the communication system 2000, as well as a radio interface 2027 for setting up and maintaining at least a wireless connection 2070 with a UE 2030 located in a coverage area (not shown in Fig. 20) served by the base station 2020. The communication interface 2026 may be configured to facilitate a connection 2060 to the host computer 2010. The connection 2060 may be direct, or it may pass through a core network (not shown in Fig. 20) of the telecommunication system and / or through one or more intermediate networks outside the telecommunication system. In the embodiment shown, the hardware 2025 of the base station 2020 further includes processing circuitry 2028, which may comprise one or more programmable processors, application-specific integrated circuits, field programmable gate arrays or combinations of these (not shown) adapted to execute instructions. The base station 2020 further has software 2021 stored internally or accessible via an external connection.

[0292] The communication system 2000 further includes the UE 2030 already referred to. Its hardware 2035 may include a radio interface 2037 configured to set up and maintain a wireless connection 2070 with a base station serving a coverage area in which the UE 2030 is currently located. The hardware 2035 of the UE 2030 further includes processing circuitry 2038, which may comprise one or more programmable processors, application-specific integrated circuits, field programmable gate arrays or combinations of these (not shown) adapted to execute instructions. The UE 2030 further comprises software 2031, which is stored in or accessible by the UE 2030 and executable by the processing circuitry 2038. The software 2031 includes a client application 2032. The client application 2032 may be operable to provide a service to a human or non-human user via the UE 2030, with the support of the host computer 2010. In the host computer 2010, an executing host application 2012 may communicate with the executing client application 2032 via the OTT connection 2050 terminating at the UE 2030 and the host computer 2010. In providing the service to the user, the client application 2032 may receive request data from the host application 2012 and provide user data in response to the request data. The OTT connection 2050 may transfer both the request data and the user data. The client application 2032 may interact with the user to generate the user data that it provides.

[0293] It is noted that the host computer 2010, base station 2020 and UE 2030 illustrated in Fig. 20 may be identical to the host computer 1930, one of the base stations 1912a, 1912b, 1912c and one of the UEs 1991, 1992 of Fig. 19, respectively. This is to say, the inner workings of these entities may be as shown in Fig. 20, and, independently, the surrounding network topology may be that of Fig. 19.

[0294] In Fig. 20, the OTT connection 2050 has been drawn abstractly to illustrate the communication between the host computer 2010 and the UE 2030 via the base station 2020, without explicit reference to any intermediary devices and the precise routing of messages via these devices. Network infrastructure may determine the routing, which it may be configured to hide from the UE 2030 or from the service provider operating the host computer 2010, or both. While the OTT connection 2050 is active, the network infrastructure may further take decisions by which it dynamically changes the routing (e.g., on the basis of load balancing consideration or reconfiguration of the network).

[0295] The wireless connection 2070 between the UE 2030 and the base station 2020 is in accordance with the teachings of the embodiments described throughout this disclosure. One or more of the various embodiments improve the performance of OTT services provided to the UE 2030 using the OTT connection 2050, in which the wireless connection 2070 forms the last segment. More precisely, the teachings of these embodiments may reduce the latency and improve the data rate and thereby provide benefits such as better responsiveness and improved QoS.

[0296] A measurement procedure may be provided for the purpose of monitoring data rate, latency, QoS and other factors on which the one or more embodiments improve. There may further be an optional network functionality for reconfiguring the OTT connection 2050 between the host computer 2010 and UE 2030, in response to variations in the measurement results. The measurement procedure and / or the network functionality for reconfiguring the OTT connection 2050 may be implemented in the software 2011 of the host computer 2010 or in the software 2031 of the UE 2030, or both. In embodiments, sensors (not shown) may be deployed in or in association with communication devices through which the OTT connection 2050 passes; the sensors may participate in the measurement procedure by supplying values of the monitored quantities exemplified above, or supplying values of other physical quantities from which software 2011, 2031 may compute or estimate the monitored quantities. The reconfiguring of the OTT connection 2050 may include message format, retransmission settings, preferred routing etc.; the reconfiguring need not affect the base station 2020, and it may be unknown or imperceptible to the base station 2020. Such procedures and functionalities may be known and practiced in the art. In certain embodiments, measurements may involve proprietary UE signaling facilitating the host computer's 2010 measurements of throughput, propagation times, latency and the like. The measurements may be implemented in that the software 2011, 2031 causes messages to be transmitted, in particular empty or "dummy" messages, using the OTT connection 2050 while it monitors propagation times, errors etc.

[0297] Fig. 21 is a flowchart illustrating a method implemented in a communication system, in accordance with one embodiment. The communication system includes a host computer, a base station and a UE which may be those described with reference to Figs. 19 and 20. For simplicity of the present disclosure, only drawing references to Fig. 21 will be included in this paragraph. In a first step 2110 of the method, the host computer provides user data. In an optional substep 2111 of the first step 2110, the host computer provides the user data by executing a host application. In a second step 2120, the host computer initiates a transmission carrying the user data to the UE. In an optional third step 2130, the base station transmits to the UE the user data which was carried in the transmission that the host computer initiated, in accordance with the teachings of the embodiments described throughout this disclosure. In an optional fourth step 2140, the UE executes a client application associated with the host application executed by the host computer.

[0298] Fig. 22 is a flowchart illustrating a method implemented in a communication system, in accordance with one embodiment. The communication system includes a host computer, a base station and a UE which may be those described with reference to Figs. 19 and 20. For simplicity of the present disclosure, only drawing references to Fig. 22 will be included in this paragraph. In a first step 2210 of the method, the host computer provides user data. In an optional substep (not shown) the host computer provides the user data by executing a host application. In a second step 2220, the host computer initiates a transmission carrying the user data to the UE. The transmission may pass via the base station, in accordance with the teachings of the embodiments described throughout this disclosure. In an optional third step 2230, the UE receives the user data carried in the transmission.

[0299] As has become apparent from above description, at least some embodiments of the technique can improve network efficiency, e.g. by using machine learning algorithms and Al, the wireless access network can optimize the selection of access points, RRM parameter values, and beamforming, leading to improved network efficiency and performance. Same or further embodiments can enhance user experience, e.g. since with Al-driven access networks, user experience can be enhanced by minimizing disruptions and maximizing overall network efficiency. The ML system can adaptively select the best access point, RRM parameters, and beamforming to provide the most suitable connectivity for the user.

[0300] Same or further embodiments can increased security. The use of a scrambler and a permutation-based architecture of the CP including the ML system adds an extra layer of security to the wireless access network. The scrambling of the APs and the asymmetric architecture make it challenging for attackers to predict the relationship between the input and output of the ML system, reducing the effectiveness of potential attacks.

[0301] Same or further embodiments can provide flexibility and adaptability. The ML system can be trained to adapt to changing network conditions and user requirements. The permutation-based architecture of the CP enables the ML system to handle multiple APs and select the best AP, RRM parameters, or beam dynamically based on real-time measurement values.

[0302] Same or further embodiments can efficiently allocate resource of the wireless access network. By applying machine learning techniques, the wireless access network can allocate wireless resources, such as transmit power, user allocation, beamforming, data rates, handover criteria, modulation schemes, and error coding schemes, more efficiently. This ensures optimal resource usage and maximizes overall network capacity.

[0303] Same or further embodiments can improve D-MIMO operations. Distributed Multiple-Input Multiple-Output (D-MIMO) operations can be optimized with the help of Al algorithms. The ML system can select the best APs, RRM parameters, and beams for D-MIMO operations, improving coverage and capacity in the wireless access network.

[0304] Same or further embodiments can be deployed faster. The permutation-based architecture of the CP and the use of configuration messages enable faster deployment and configuration of the wireless access network. The CP can quickly determine the suitable APs and the core node can adapt the scrambler and / or the ML system accordingly, reducing overall network setup time.

[0305] Many advantages of the present invention will be fully understood from the foregoing description, and it will be apparent that various changes may be made in the form, construction and arrangement of the units and devices without departing from the scope of the invention and / or without sacrificing all of its advantages. Since the invention can be varied in many ways, it will be recognized that the invention should be limited only by the scope of the following claims.

Claims

Claims1. A method (300) performed by a wireless access network (500) comprising access points, APs (502), the method (300) comprising: obtaining (302) measurement values indicative of wireless links between a wireless device, WD (504; 1991; 1992; 2030), and multiple APs (502) distributed in the wireless access network (500); and serving (304) or initiating (304) to serve the WD (504; 1991; 1992; 2030) using at least one AP (502; 130) among the APs (502), wherein the at least one AP (502; 130) is determined using a central processor, CP (100), comprising a scrambler (110) and a machine learning system, ML system (103), wherein the scrambler (110) applies the obtained (302) measurement values to an input layer (101) of the ML system (103) according to a permutation of the multiple APs (502), and an architecture of the ML system (103) is asymmetric with respect to the permutation of the multiple APs (502).

2. The method (300) of claim 1, wherein the at least one AP (502; 130) is determined from a set of APs (502) of the wireless access network (500), and wherein the obtained measurement values relate to the multiple APs (502) which are a proper subset of the set of APs (502).

3. The method (300) of claim 1 or 2, wherein each of the obtained (302) measurement values is indicative of one of the wireless links using one AP (502) of the multiple APs (502) and one radio resource management, RRM, parameter value of an RRM parameter of the respective one AP (502).

4. The method (300) of any one of claims 1 to 3, wherein the WD (504; 1991; 1992; 2030) is served (304) according to an RRM parameter of the at least one AP (502), and wherein an RRM parameter value (130) of the RRM parameter is determined using the CP (100).

5. The method (300) of claim 4, wherein the RRM parameter value (130) is determined from a set of RRM parameter values of the at least one AP (502), and wherein the one or more measurement values obtained (302) for said at least one AP (502) relate to a proper subset of the set of RRM parameter values.

6. The method (300) of any one of claims 3 to 5, wherein the RRM parameter comprises at least one of transmit power, user allocation, beamforming, data rate, handover criterion, modulation scheme, and error coding scheme, and / or wherein the RRM parameter value (130) is indicative of at least one of a transmit power level, wireless communication resources allocated to the WD (504; 1991; 1992; 2030), a beam at the at least one AP (502), a data rate, a handover criterion, a modulation scheme, and an error coding scheme.

7. The method (300) of any one of claims 1 to 6, wherein the wireless links correspond to beams at the multiple APs (502), and / or wherein each of the obtained (302) measurement values is indicative of one of the wireless links using one AP (502) of the multiple APs (502) and one beam (130) of the respective one AP (502).

8. The method (300) of any one of claims 1 to 7, wherein the WD (504; 1991; 1992; 2030) is served (304) using beamforming at the at least one AP (502), and wherein a beam (130) of the at least one AP (502) is determined using the CP (100).

9. The method (300) of claim 8, wherein the beam (130) is determined from a set of beams of the at least one AP (502), and wherein the one or more measurement values obtained (302) for the at least one AP (502) relate to a proper subset of the set of beams.

10. The method (300) of any one of claims 1 to 9, wherein the ML system (103) determines at least one of: a beam (130) for the serving (302) of the WD (504; 1991; 1992; 2030); a combination of the at least one AP (502) and a beam (130) for the serving (302) of the WD (504; 1991; 1992; 2030); and a downlink transmit power (130) for the serving (302) of the WD (504; 1991; 1992; 2030).

11. The method (300) of any one of claims 1 to 10, wherein the measurement values are obtained (302) from the multiple APs (502), and / or wherein the measurement values are indicative of the wireless links measured at the multiple APs (502) with respect to the WD (504; 1991; 1992; 2030), and / or wherein the measurement values are based on uplink reference signals received at the multiple APs (502) from the WD (504; 1991; 1992; 2030).

12. The method (300) of any one of claims 1 to 11, wherein the measurement values are obtained (302) from the WD (504; 1991; 1992; 2030), and / or wherein the measurement values are indicative of the wireless links measured at the WD (504; 1991; 1992; 2030) with respect to the multiple APs (502), and / or wherein the measurement values are based on downlink reference signals received from the multiple APs (502) at the WD (504; 1991; 1992; 2030), and / or wherein the obtaining (302) of the measurement values comprises receiving measurement reports comprising the measurement values, each of the measurement reports relating exclusively to one of the multiple APs (502).

13. The method (300) of any one of claims 1 to 12, wherein the at least one AP (502) comprises two or more APs (502) determined for serving (304) the WD (504; 1991; 1992; 2030) in dual connectivity, DC, or multi-connectivity, MC.

14. The method (300) of any one of claims 1 to 13, wherein the multiple APs (502) or the set of APs (502) are distributed across a site of coherent coverage of the wireless access network (500).

15. The method (300) of any one of claims 1 to 14, wherein the wireless access network (500) serves (304) the WD (504; 1991; 1992; 2030) according to a distributed multiple-input multiple-output, D-MIMO, operation of the determined at least one AP (502) or the set of APs (502), and / or wherein the multiple APs (502) or the set of APs (502) belong to a D-MIMO site of the wireless access network (500), and / or wherein the obtained (302) measurement values are indicative of a link quality of at least one beam from each of the multiple APs (502) belonging to a D-MIMO site of a D-MIMO operation of the wireless access network (500).

16. The method (300) of any one of claims 1 to 15, wherein the measurement values are indicative of a link quality of each of the wireless links, and / or wherein the measurement values comprise values of a reference signal received power, RSRP, or values of a reference signal received quality, RSRQ, or received signal strength indicators, RSSIs.

17. The method (300) of any one of claims 1 to 16, wherein the measurement values of the wireless links are based on reference signals transmitted from the multiple APs (502) on the respective wireless links, optionally wherein the reference signals comprise synchronization signals periodically broadcasted in the downlink on the beams from the plurality of APs (502) and / or primary synchronization signals, PSSs, and / or secondary synchronization signals, SSSs, and / or demodulation reference signals, DM-RS.

18. The method (300) of any one of claims 1 to 17, wherein an order of the multiple APs (502) is scrambled according to the same permutation during both training of the ML system (103) and the determining of the at least one AP (502) or the RRM parameter value or the beam using the ML system (103).

19. The method (300) of any one of claims 1 to 18, wherein the WD (504; 1991; 1992; 2030) is wirelessly connected to the wireless access network (500), and / or wherein the measurement values are measured at the radio device, optionally measured for wireless access to the wireless access network (500) or measured for mobility of the WD (504; 1991; 1992; 2030) in a D-MIMO site of the wireless access network (500) or measured during an idle or inactive state of the WD (504; 1991; 1992; 2030) for early reporting enabling faster setup of the D- MIMO operation for the WD (504; 1991; 1992; 2030).

20. The method (300) of any one of claims 1 to 19, wherein the method (300) is performed by one of the multiple APs (502) or one of the set of APs (502) of the wireless access network (500), or wherein the method (300) is performed by the CP (100) of the wireless access network (500), optionally wherein the CP (100) is hosting the ML system (103) and / or the CP (100) is controlling the D-MIMO operation and / or the CP (100) is connected to each of the multiple APs (502) or each of the set of APs (502).

21. The method (300) of any one of claims 1 to 20, wherein the applying of the obtained (302) measurement values comprises permuting the order of the multiple APs (502) according to the permutation of the scrambler (110) at the input layer (101) of the ML system (103) without changing an order of RRM parameter values or beams within a feature group of the respective AP (502) at the input layer (101) of the ML system (103).

22. The method (300) of any one of claims 1 to 21, wherein the wireless access network (500) comprises multiple D-MIMO sites each comprising a CP (100) and multiple APs (502) for serving a WD (504; 1991; 1992; 2030) in a D-MIMO operation, and wherein different permutations of the multiple APs (502) are used for the applying of the measurement values obtained (302) at the different D- MIMO sites.

23. The method (300) of claim 22, wherein the permutations are generated independently at each of the D-MIMO sites based on a source of true randomness other than a pseudorandom number generator and / or wherein the permutations of the D-MIMO sites are mutually compared by exchanging between the CPs (100) results of a hash function or a public key generator applied to an identifier of each of the permutations and / or wherein at least one of the permutations is regenerated if the generated permutations are equal.

24. The method (300) of any one of claims 1 to 23, further comprising: receiving a configuration message (210) indicative of the permutation, optionally wherein the configuration message (210) is encrypted.

25. The method (300) of any one of claims 1 to 24, wherein the architecture of the ML system (103) comprises a convolutional filter or a convolutional neural network, and / or wherein the input layer of the ML system (103) comprises a convolutional filter or a convolutional neural network, and / or wherein a size of a kernel of the convolutional filter or the convolutional neural network or the input layer of the ML system (103) is multiple times smaller than the number of the multiple APs (502).

26. A method (400) performed by a core node (200; 1800; 1914; 1920; 1930; 2010) configuring a wireless access network (500) comprising access points, APs (502), the method (400) comprising: determining (402) a permutation of multiple APs (502) distributed in the wireless access network (500), wherein the permutation is applied prior to a machine learning system, ML system (103), that processes measurement values indicative of wireless links between a wireless device, WD (504; 1991; 1992; 2030), and the multiple APs (502), wherein an architecture of the ML system (103) is asymmetric with respect to the permutation of the multiple APs (502); and sending (404) a configuration message (210) indicative of the determined (402) permutation to a central processor, CP (100), configured to determine at least one AP (502), wherein a scrambler (110) of the CP (100) applies the measurement values to an input layer (101) of the ML system (103) of the CP (100) according to the permutation.

27. The method (400) of claim 26, wherein the determining (402) of the permutation comprises duplicating or replicating the ML system (103) of the CP (100) at the core node (200; 1800; 1914; 1920; 1930; 2010).

28. The method (400) of claim 1 , wherein the permutation is determined (404) dependent on the measurement values, optionally RSRP values, obtained from the wireless access network (500) or the WD (504; 1991; 1992; 2030).

29. The method (400) of claim 1 or 28, wherein the permutation is determined by selecting the permutation from a plurality of candidate permutations by measuring a performance measure of the duplicated or replicated ML system (103) in combination with each of the permutation candidates, wherein the duplicated or replicated ML system (103) is retrained for each of the candidate permutations.

30. The method (400) of any one of claims 27 to 29, wherein the configuration message (210) sent (404) to the CP (100) is further indicative of the ML system (103) retrained for the determine (402) permutation.

31. The method (400) of any one of claims 27 to 30, wherein the wireless access network (500) comprises at least two sites each comprising multiple APs (502)and a CP (100)7and wherein the permutation is determined (402) for each of the at least two sites, and wherein configuration messages (210), each indicative of a respective one of the at least two permutations, are sent (404) to the respective CPs (100) of the at least two sites.

32. The method (400) of claim 31, wherein different permutations are determined for different sites of the wireless access network (500).

33. The method (400) of any one of claims 26 to 32, wherein the sent (404) configuration message (210) is encrypted or the sent (404) configuration messages (210) are encrypted.

34. The method (400) of any one of claims 26 to 33, further comprising the features or steps of any one of claims 2 to 25, or any feature or step corresponding thereto.

35. A computer program product comprising program code portions for performing the steps of any one of the claims 1 to 25 or 26 to 34 when the computer program product is executed on one or more computing devices (1704; 1804), optionally stored on a computer-readable recording medium (1706; 1806).

36. A node (100; 1700; 502; 1911; 1912; 2020) of a wireless access network (500) comprising access points, APs (502), the node (100; 1700; 502; 1911; 1912; 2020) comprising memory operable to store instructions and processing circuitry operable to execute the instructions, such that the node (100; 1700; 502; 1911; 1912; 2020) is operable to: obtain (302) measurement values indicative of wireless links between a wireless device, WD (504; 1991; 1992; 2030), and multiple APs (502) distributed in the wireless access network (500); and serve (304) or initiating (304) to serve the WD (504; 1991; 1992; 2030) using at least one AP (502) among the APs (502), wherein the at least one AP (502) is determined using a central processor, CP (100), comprising a scrambler (110) and a machine learning system, ML system (103), wherein the scrambler (110) applies the obtained (302) measurement values to an input layer (101) of the ML system (103) according to a permutation of the multiple APs (502), and an architectureof the ML system (103) is asymmetric with respect to the permutation of the multiple APs (502).

37. The node (100; 1700; 502; 1911; 1912; 2020) of claim 36, further operable to perform the steps of any one of claims 2 to 25.

38. A node (100; 1700; 502; 1911; 1912; 2020) of a wireless access network (500) comprising access points, APs (502), the node (100; 1700; 502; 1911; 1912; 2020) being configured to: obtain (302) measurement values indicative of wireless links between a wireless device, WD (504; 1991; 1992; 2030), and multiple APs (502) distributed in the wireless access network (500); and serve (304) or initiating (304) to serve the WD (504; 1991; 1992; 2030) using at least one AP (502) among the APs (502), wherein the at least one AP (502) is determined using a central processor, CP (100), comprising a scrambler (110) and a machine learning system, ML system (103), wherein the scrambler (110) applies the obtained (302) measurement values to an input layer (101) of the ML system (103) according to a permutation of the multiple APs (502), and an architecture of the ML system (103) is asymmetric with respect to the permutation of the multiple APs (502).

39. The node (100; 1700; 502; 1911; 1912; 2020) of claim 38, further configured to perform the steps of any one of claims 2 to 25.

40. A core node (200; 1800; 1914; 1920; 1930; 2010) for configuring a wireless access network (500) comprising access points, APs (502), the core node (200; 1800; 1914; 1920; 1930; 2010) comprising memory operable to store instructions and processing circuitry operable to execute the instructions, such that the core node (200; 1800; 1914; 1920; 1930; 2010) is operable to: determine (402) a permutation of multiple APs (502) distributed in the wireless access network (500), wherein the permutation is applied prior to a machine learning system, ML system (103), that processes measurement values indicative of wireless links between a wireless device, WD (504; 1991; 1992; 2030), and the multiple APs (502), wherein an architecture of the ML system (103) is asymmetric with respect to the permutation of the multiple APs (502); andsend (404) a configuration message (210) indicative of the permutation to a central processor, CP (100), configured to determine at least one AP (502), wherein a scrambler (110) of the CP (100) applies the measurement values to an input layer (101) of the ML system (103) of the CP (100) according to the permutation.

41. The core node (200; 1800; 1914; 1920; 1930; 2010) of claim 40, further operable to perform the steps of any one of claims 27 to 34.

42. A core node (200; 1800; 1914; 1920; 1930; 2010) for configuring a wireless access network (500) comprising access points, APs (502), the core node (200; 1800; 1914; 1920; 1930; 2010) being configured to: determine (402) a permutation of multiple APs (502) distributed in the wireless access network (500), wherein the permutation is applied prior to a machine learning system, ML system (103), that processes measurement values indicative of wireless links between a wireless device, WD (504; 1991; 1992; 2030), and the multiple APs (502), wherein an architecture of the ML system (103) is asymmetric with respect to the permutation of the multiple APs (502); and send (404) a configuration message (210) indicative of the permutation to a central processor, CP (100), configured to determine at least one AP (502), wherein a scrambler (110) of the CP (100) applies the measurement values to an input layer (101) of the ML system (103) of the CP (100) according to the permutation.

43. The core node (200; 1800; 1914; 1920; 1930; 2010) of claim 42, further configured to perform the steps of any one of claims 27 to 34.

44. A communication system (1400; 1900; 1910; 2000) including a host computer (1330; 1410) comprising: processing circuitry (1418) configured to provide user data; anda communication interface (1416) configured to forward user data to a cellular or ad hoc or D-MIMO radio network (1400; 1900; 1910; 2000) for transmission to a user equipment, UE (504; 1991; 1992; 2030) wherein radio network (1400; 1900; 1910; 2000) comprises a radio interface (1702; 2037) and processing circuitry (1704; 2038), the processing circuitry (1704; 2038) of the radio network (1400; 1900; 1910; 2000) being configured to execute the steps of any one of claims 1 to 24.

45. The communication system (1400; 1900; 1910; 2000) of claim 44, further including a UE (504; 1991; 1992; 2030).

46. The communication system (1400; 1900; 1910; 2000) of claim 44 or 45, wherein the radio network (1400; 1900; 1910; 2000) further comprises a base station (100; 1700; 502; 1911; 1912; 2020), or a radio device (504; 1991; 1992; 2030) functioning as a gateway, which is configured to communicate with the UE (504; 1991; 1992; 2030).

47. The communication system (1400; 1900; 1910; 2000) of claim 46, wherein the base station (100; 1700; 502; 1911; 1912; 2020), or the radio device (504; 1991; 1992; 2030) functioning as a gateway, comprises processing circuitry (1704; 2028), which is configured to execute the steps of any one of claims 1 to 24.

48. The communication system (1400; 1900; 1910; 2000) of any one of claims 44 to 47, wherein: the processing circuitry (2018) of the host computer (1930; 2010) is configured to execute a host application (2012), thereby providing the user data; and the processing circuitry (2038) of the UE (504; 1991; 1992; 2030) is configured to execute a client application (2032) associated with the host application (2012).

Images