Method and system for authentication of RF device

Abstract

The method for authenticating a first device (10) includes: A) receiving an RF signal from the first device (10); B) obtaining feature data from the received RF signals (SG1, SG2, …); C) adapting the obtained feature data to a model (M) for authenticating the first device (10); D) storing the adapted model (M1) in memory in a second device (30); and by means of the second device (30), E) receiving a new RF signal from a device (T) to be authenticated; F) obtaining feature data from the new RF signal; and G) determining whether the device (T) to be authenticated is the first device (10) by transmitting the feature data obtained from the new RF signal (SG(t)) as input to the adapted model (M1). The feature data obtained from an RF signal includes the value of only one of a co-phase component I(t) and a quadrature component Q(t) of the RF signal over time.

Description

Technical Field

[0001] This disclosure relates to the authentication of a first RF transmitter device by means of a second radio frequency (RF) device. Prior Technology

[0002] Today, many devices, from mobile phones to IoT sensors / devices, use radio spectrum communication. Generally, devices carry an ID (in other words, an identifier) ​​and implement security measures such as encryption, identification, and / or authentication to ensure the confidentiality and / or authenticity of RF communications. In the IoT field, devices are often low-cost and based on hardware with low security levels. Generally, they have IDs, but their hardware security is unreliable.

[0003] The main threat to RF communication systems is that a device can be cloned by acquiring all of its memory and copying it onto a counterfeit device that appears to be the genuine device at the hardware level. The counterfeit device can then send (multiple) messages just like the genuine device. It is difficult to distinguish between the two devices (the genuine and the counterfeit) at the time.

[0004] Authentication is a crucial step in establishing a communication system for trusted devices.

[0005] For radio communication, RF devices have a chain of analog components, such as oscillators, inductors, capacitors, etc. Analog components are not perfect, and these defects affect the form of the electromagnetic spectrum transmitted by the RF device via radio waves. For example, due to an imperfect oscillator, the carrier frequency may have a shift. This defect (carrier frequency shift, or "CFO") does not affect the operation of the RF device, but it can be observed in its transmission spectrum. This defect establishes a fingerprint of the RF device within its RF transmission. Because hardware defects are virtually impossible to counterfeit, the resulting fingerprint can be used to identify and / or authenticate RF transmitter devices. The use of this fingerprint to identify and / or authenticate RF transmitter devices is well known.

[0006] For example, US 10,693,576 B2 discloses a fingerprint authentication method for RF devices based on the CFO (Carrier Frequency Offset) and operating and / or environmental conditions (such as temperature, relative humidity, etc.) used as RF fingerprints. More precisely, it calculates and uses the CFO value derived from the RF transmission of the RF device as the RF fingerprint of the RF device. In practice, small defects in the analog components of the RF device cause small variations in its RF transmission. The method disclosed in US 10,693,576 B2 includes the following steps: A plurality of training RF transmissions are received from an RF device via an RF receiver, each of which is temporarily associated with operating parameters and environmental parameters of the RF receiver and the RF device. During the training phase, the machine learning classifier is trained at least in part based on a training set containing: (i) The carrier frequency offset (CFO) values ​​calculated for each of the training RF transmissions, and (ii) Tags associated with these operating parameters and these environmental parameters; and During the inference phase, the machine learning classifier is applied to determine whether one or more execution-time RF transmissions originate from the RF device.

[0007] However, the above method has some drawbacks. In practice, fingerprint authentication as disclosed in US 10,693,576 B2 requires computational components to calculate the CFO value from the RF transmission. Furthermore, the surrounding environment and / or operating conditions need to be determined (which requires some sensors in the RF device) to more accurately authenticate the RF transmitter device.

[0008] Therefore, there is a need to promote the authentication of RF transmissions based on RF devices. Summary of the Invention

[0009] This disclosure relates to a computer implementation method for authentication of a first device, the computer implementation method comprising the following steps: A) Receive a plurality of RF signals from the first device; B) Obtain characteristic data from the received RF signals; C) Using the acquired feature data, a model is adapted for authenticating the first device; D) Store the adapted model in the memory of a second device; E) Receive a new RF signal from a device to be authenticated via the second device; F) Obtain characteristic data from the new RF signal using the second device; and G) Using the second device, the characteristic data obtained from the new RF signal is transmitted as input to the adapted model to determine whether the device to be authenticated is the first device; The characteristic data obtained from an RF signal includes the value of only one of the co-phase component I(t) of the RF signal over time and the quadrature component Q(t) of the RF signal over time.

[0010] The feature data used to adapt the model via received RF signals includes I-data or Q-data. Once the model has been adapted using the I-data or Q-data (e.g., for training), it can be stored in a second RF device. The second RF device is then paired with the first RF device. It is anticipated that the two paired RF devices will communicate with each other. For example, the first RF device is an IoT device or sensor, and the second device is an IoT gateway. When the first RF device wants to connect to and / or communicate with the second RF device, the latter performs steps E) to G) as defined above to allow (or disallow) the first RF device to access the second RF device and / or further online services.

[0011] The method defined above for authenticating the first RF device based on the RF signals transmitted by the first RF device facilitates authentication, improves its performance, and reduces computational workload. Authentication is based on I-data and / or Q-data obtained from the RF signals received from the first RF device, which facilitates the implementation of this method. The I-data or Q-data is used as a fingerprint of the first RF device. Authentication is highly accurate. Authentication is reliable and effective without requiring environmental and / or operating conditions. This method allows for well-designed authentication schemes while reducing manufacturing costs and computational workload.

[0012] Furthermore, training or adapting the model using only I or Q data allows for a significant reduction in training time. Next, in the step of determining whether the device to be certified is the same as the first device using the adapted or trained model via a second device, there is no significant difference in certification accuracy between using only I or Q data and using both I and Q data. This allows for a reduction in the amount of data processed by the model.

[0013] The model may be a machine learning model, and the step of adapting the model may include a step of using the feature data as input training data to train the machine learning model.

[0014] In one embodiment, the machine learning model is trained in a semi-supervised manner using input training data that exclusively includes feature data obtained from RF signals transmitted from the first device.

[0015] In one embodiment, the model is an anomaly detection model. In data analysis, anomaly detection (also known as outlier detection) identifies rare items, events, or observations that are significantly different from the majority of data, raising suspicion. Anomaly detection is particularly well-suited to the context of authentication and counterfeit detection. It allows newly incoming RF signals, after being converted into characteristic data, to be labeled as either anomalous or normal RF signals. In other words, anomaly detection allows the detection of whether a newly incoming RF signal is a normal RF signal sent from a genuine RF device or an anomalous RF signal (sent from a counterfeit RF device).

[0016] After the adapted model is stored in memory, it can be updated by the second device based on feature data obtained from a new RF signal received from the first device. Therefore, after being stored in memory in the second RF device, the adapted model can be updated using new feature data from new RF signals(s) received from the first device. This allows the model to adapt to various environmental conditions that may not be included in the initial model stored in memory in the second RF device.

[0017] Optionally, the method further includes a step of first normalizing the feature data (in other words, the I data or the Q data) in at least one of steps C) and G) and then providing the normalized feature data to the model. The main advantage of data normalization is that it makes the data less dependent on the strength of the signal received by the second RF device. In fact, the strength of the received signal affects the absolute value of the feature data: the smaller the received signal strength, the closer the received feature value is to 0. By forcing all values ​​to be on the same scale through normalization, the model becomes less sensitive to the strength of the received signal.

[0018] Optionally, the method further includes a step of first filtering the feature data in at least one of the steps C) and G) and then providing the filtered feature data to the model.

[0019] Some received RF signals may require processing or removal. For example, the duration of an RF signal may be unusually short compared to the average duration of an RF signal. Removing or processing the RF signal allows for improvements in the efficiency and accuracy of the method. Optionally, downsampling can be performed. This essentially involves removing samples (and typically also applying a low-pass filter). The target system has fewer samples to process.

[0020] Steps A) through C) can be performed by the second device or, alternatively, by a third device different from the second device.

[0021] In one embodiment, steps A) to C) are also performed on a plurality of RF signals transmitted by the second device to adapt to a model for authenticating the second device, and steps D) to G) are also performed by the first device to authenticate the second device using the adapted model, such that the first device and the second device perform a mutual authentication.

[0022] In one embodiment, steps A) to C) are performed on a plurality of first devices, and a plurality of adapted model systems corresponding to the plurality of first devices are stored in the memory of the second device.

[0023] The method may further include A step of establishing communication between the device to be authenticated and the second device, wherein the device to be authenticated transmits an identifier to the second device; and The second device selects an adapted model corresponding to the device to be authenticated from among the plurality of adapted models stored in memory based on the identifier of the device to be authenticated, and the selected adapted model is used in step G).

[0024] This disclosure also relates to a system for authentication of a first device, the system comprising one of the following A second device configured to perform steps A) to G) of the previously defined method, Or a second device configured to perform steps D) to G) of the previously defined method, and a third device configured to perform steps A) to C) of the previously defined method.

[0025] This disclosure also relates to a computer program containing instructions that, when executed by a computer, cause the computer to perform the steps of a previously defined computer implementation method. Simple Explanation of the Diagram

[0026] Other features, objects, and advantages of this disclosure will become more apparent from the detailed description of the non-limiting embodiments made with reference to the accompanying drawings. Figure 1 is a block diagram of a system for authenticating a first RF device based on an RF signal originating from the first RF device, according to one embodiment. Figure 2 is a block diagram of a second RF device according to an embodiment of the system of Figure 1. Figure 3 is a block diagram of a third RF device according to an embodiment of the system of Figure 1. Figure 4 is a flowchart of the initial (or training) phase of a computer implementation method for authentication of a first RF device according to an embodiment. Figure 5 is a flowchart of the operation mode (or inference phase) of a computer implementation method for authenticating a first RF device according to an embodiment. Implementation

[0027] The computer implementation method disclosed herein can be used to identify and / or authenticate a wireless or RF (radio frequency) first device 10. The authentication of the first RF device 10 is based on the RF signals transmitted by the first RF device 10, and more precisely, on the characteristic data of the RF signals from the first RF device 10.

[0028] The computer implementation method includes adapting (or training) an initial phase (program) for a model used for authentication of the first device 10, and using the adapted model by the second RF device 30 to authenticate the operating phase of the first device 10. The initial phase may be executed by the second RF device 30 or alternatively by a third RF device 20 different from the second RF device 30.

[0029] The first RF device 10 may be an RF transmitter or an RF transceiver. The second RF device 30 may be a receiver or a transceiver. The third RF device 20 may be a receiver or a transceiver. The first RF device 10 and the second RF device 30 are paired or will be paired with each other. As an illustrative and non-limiting example, the first RF device 10 is an IoT device or a sensor, and the second RF device 30 is an IoT gateway.

[0030] In the initial (in other words: training or learning) phase, the second RF device 30 (or alternatively, the third RF device 20) receives a plurality of RF signals SGi(t) transmitted from the first RF device 10, where i = 1, 2, 3, …, or RF messages. In one embodiment, the RF signals transmitted by the first RF device 10 are repetitions of the same message having the same content and the same device and / or agreed-upon specific parameters. In this case, the RF device 10 repeatedly sends the same message. This could be a message intended to be sent by the RF device 10 later in an operational mode (in other words: during the use of the RF device 10 in a real-world environment). In another embodiment, different messages may be sent. Specifically, if the RF device 10 intends to send different messages during its lifetime (in an operational mode), these messages can be used to effectively train the model. More precisely, the feature data obtained from the received RF signals SGi(t) is then used to adapt (in other words: learn or train) the model M responsible for (later) performing the task of authenticating the first RF device 10. The number of messages sent to obtain a high-performance model for authentication may depend on the first device 10, the content of the messages, and / or other parameters. For example, hundreds of messages may be sent to correctly adapt the model M to the first RF device 10. Once adapted to feature data obtained from multiple RF signals originating from the RF device 10, the model M is stored and the second RF device 30 stores it in memory. Then, in operating mode (in other words: in the inference phase), the second RF device 30 may authenticate the first (known) RF device 10 by transmitting (providing) one or more new RF signals transmitted from the first RF device 10 to the second RF device 30 as input to the model M stored in memory.

[0031] In one embodiment, the characteristic data obtained from any RF signal (or message) includes I and Q data of the RF signal over time; in other words, a series of values ​​for the in-phase component I(t) and the quadrature component Q(t) of the received RF signal SGi(t) over time. The I and Q data are obtained from the received RF signal SGi(t) over time by components of the radio receiver and / or by the quadrature demodulator (also referred to as I / Q demodulator) of the second device 30 or the third device 20.

[0032] In other embodiments, the feature data obtained from the received RF signal (or message) and to be provided to model M (in the initial phase) or M1 (in the operating mode) includes only I data (in other words, a series of values ​​of the in-phase component I(t) of the received RF signal over time) or only Q data (in other words, a series of values ​​of the quadrature component Q(t) of the received RF signal over time).

[0033] Characteristic data obtained from any RF signal originating from the RF device will be used as the fingerprint of that RF device.

[0034] In one embodiment, the model is an anomaly detection model. Anomaly detection (AD) (also known as outlier detection) aims to mark newly incoming points as either anomalous or normal. In the context of fingerprint authentication, the goal is to detect whether newly incoming messages (in other words, new received RF signals) are normal messages sent by a known transmitter device or anomalous messages sent by another device (such as a cloning device). Therefore, the advantage of using an anomaly detection model lies in the interpretability and understanding of the strategy.

[0035] In one embodiment, model M is a machine learning model. It is trained to authenticate the first RF device 10 using feature data (in other words: I data and Q data, or only I data, or only Q data) obtained from each of the plurality of RF signals received from the first RF device 10. Therefore, model M learns the specificity of the first RF device 10.

[0036] Alternatively, model M can be a system design model, a spectrum model, or a model based on information theory.

[0037] In the operating mode (in other words: in the inference phase), the second RF device 30 is responsible for authenticating the first RF device 10 using a previously adapted or trained stored model M1.

[0038] When the second RF device 30 receives a new RF signal from the RF device T to be authenticated, it obtains feature data (in other words: I data and Q data, or alternatively only I data or Q data) from the received RF signal and transmits (provides) the obtained feature data as input data to the model M1 stored in memory to determine whether the RF device T to be authenticated is the first RF device 10.

[0039] During the initial (training) phase and the operational (inference) phase, feature data (I-data and Q-data, or I-data or Q-data only) may optionally be preprocessed before being provided as input to the model. Preprocessing may include normalization and / or filtering of the data by removing and / or processing it, as explained in more detail later.

[0040] Figure 1 is a block diagram of a system 100 for authenticating the first RF device 10 based on an RF signal originating from the first RF device 10, according to one embodiment.

[0041] System 100 includes a second RF device 30 and optionally a third RF device 20. The role of the second RF device 30 (or alternatively, the third RF device 20) is to adapt the model M1 to the RF received signals transmitted by the first RF device 10. The second RF device 30 stores the model M1 adapted to the first RF device 10 in memory. Then, the second RF device 30 can authenticate the first RF device 10 using the model M1 and one or more RF signals transmitted from the first RF device 10 and provided as input to the model M1.

[0042] The second RF device 30 illustrated in Figure 2 includes an antenna 300, a radio receiver 301 with an IQ demodulator 302, a processor or CPU (central processing unit) 303, RAM 304, a storage unit 305, an adapter or training component 306 for training (adapting) the model and optionally updating the adapted model, and an authentication component 310.

[0043] As is well known to those skilled in the art, radio receiver 301 receives radio waves via antenna 300 and converts the information carried by the radio waves into a usable form. It includes various components (mixers, filters, amplifiers, modulators / demodulators, detectors, etc.) that can be implemented in hardware and / or software. In one embodiment, the radio receiver includes components implemented in software as a software-defined radio or SDR chip. Alternatively, the radio receiver can be implemented in hardware or by a hybrid of hardware and software.

[0044] Advantageously, the second RF device 30 introduces low jitter (bias) into the received RF signal. In fact, since jitter or bias also depends on the analog components, the second RF device 30 may introduce it into the received RF signal. This bias can be significantly reduced by using high-quality components. This is particularly noteworthy in embodiments of system 100 that include both the second RF device 30 and the third RF device 20. As will be explained later, in such embodiments, the third RF device 20 is responsible for adapting (training) the model M to the RF received signals transmitted by the first RF device 10 (with which it is trained), and the second RF device 30 then uses the adapted model. However, in embodiments of system 100 that include the second RF device 30, which is responsible for adapting (training) the model on the first RF device 10 and then using the trained model to certify the first RF device 10, the bias introduced by the second RF device 30 is not significant. In this case (when the same receiver is used for training and operation), the use of high-quality components is not required.

[0045] In this disclosure, the radio receiver 301 includes an IQ demodulator 302 (also referred to as a quadrature demodulator) responsible for converting an RF input signal into two baseband signals, namely, an in-phase signal (component) I(t) and a quadrature signal (component) Q(t) converted into digital data by sampling. For example, if a 1-second RF input signal is acquired and the sampling rate is 100 samples per second, the IQ demodulator converts the RF input signal into 100 values ​​of the I(t) component and 100 values ​​of the Q(t) component. These values ​​can be represented in complex vector form, for example, [0.7+0.5j; 0.2+0.3j; …]. In this example, the values ​​generated by the IQ demodulator are contained within (-1.0, 1.0). Describing the first sample as "0.7+0.5j" means that the I value is "0.7" and the Q value is "0.5". A complex vector can be converted into two vectors, one containing I values ​​and the other containing Q values.

[0046] Other components of the radio receiver 301 are well known to those skilled in the art and will not be described in further detail in this disclosure.

[0047] Storage unit (memory) 305 stores one or more models for authentication of one or more RF devices. In one embodiment, storage unit 305 stores a model M to be adapted (in other words: learned or trained) on RF signals originating from the first RF device. Storage unit 205 may also store one or more adapted (trained) models M1, M2, M3, ...

[0048] The model can be adapted to a plurality of first RF devices (different RF devices). In this case, the second RF device 30 is responsible for identifying a group of first RF devices (on which the model is adapted) rather than identifying a single first device.

[0049] Alternatively, a model can be trained to recognize a specific first RF device 10. In this case, multiple different models are trained to recognize multiple first RF devices 10 (one model for one first RF device). In practice, training multiple models (one model for one first RF device) is better at rejecting counterfeit devices than adapting only one model.

[0050] CPU 303 controls the operation of the second RF device 30. RAM 304 is the working memory of CPU 303.

[0051] The training (adaptation) component 306 is a software component running on the CPU 303 to adapt or train a model stored in the storage unit 305 on an RF input signal originating from the corresponding first RF device (e.g., first RF device 10).

[0052] Optionally, the second RF device 30 has a processing component 309 including a normalization component 307 and / or a filtering component 308 (with one or more filters).

[0053] The normalization component 307 is a software component running on the CPU 303. It is responsible for normalizing the feature data (in other words: I data and Q data, or, alternatively, only I data or Q data) obtained from the RF input signal received from the radio receiver 301. Normalization enforces that all values ​​have the same scale. It can use standardization techniques. For example, removing the mean from each value and dividing the result by the standard deviation of the value scales the values ​​to a range of (-1.0; 1.0). For example, if the received I values ​​are between -0.1 and 0.6, normalization transforms them into values ​​with a mean of zero and a standard deviation of 1 in the range (-1.0; 1.0). When the feature data includes both I and Q data, normalization can be applied to the I values ​​and the Q values ​​separately. When using the previous example, the I vector is [0.7; 0.2; …] and the Q vector is [0.5; 0.3; …]. Because the I and Q values ​​may be uncorrelated, their individual distributions may differ. For example, for 100 values ​​obtained from a 1-second RF input signal, the I value can be centered at approximately 0.5 and the Q value at approximately 0.7. Therefore, the I and Q values ​​are processed into two independent (separate) vectors that must be normallyized independently. Normalization primarily allows the I and Q values ​​to be insensitive to the strength of the signal received by the receiver (here, receiver device 20). For example, distance affects the absolute range of the I and Q values: the greater the distance, the closer the received I and Q values ​​are to 0. Therefore, normalizing the I and Q vectors makes the model less sensitive to the strength of the received signal by forcing all values ​​to be on the same scale.

[0054] One or more filters or filtering components 308 are responsible for filtering the feature data (in other words: I data and Q data, or, alternatively, only I data or Q data) before providing the feature data to the training component 306. In one embodiment, the filtering component 308 is a software component running on the CPU 303. Alternatively, it may be hardware (e.g., a bandpass filter to isolate the signal of interest and avoid interference from other signals) or a mixture of software and hardware. It is responsible for filtering the feature data obtained from the RF input signal received from the radio receiver 301 by processing or removing data. For example, some RF signals received by the second RF device 30 have an anomalous duration that is very short compared to the average duration of the RF signal. Data obtained from such very short RF signals can be removed. In other cases, the filtering component 308 includes one or more of the following elements for processing the data: a noise canceller, a noise reducer, a bandpass filter, and an interference detector.

[0055] The authentication component 310 is responsible for authenticating the first RF device 10 based on one or more RF signals originating from the first RF device 10, received by the radio receiver 301 and applied (in other words: provided as input) to the stored model M1, as will be described in more detail later.

[0056] In one embodiment, the system includes a third RF device 20 (instead of the second RF device 30) illustrated in FIG3, responsible for performing an adaptation or training procedure. The third RF device 20 includes an antenna 200, a radio receiver 201, a processor or CPU (central processing unit) 203, RAM 204, a storage unit 205, and adaptation or training components 206. Advantageously, the third RF device 20 introduces low jitter (bias) into the received RF signal. In practice, because jitter or bias also depends on the analog components, the third RF device 20 may introduce it into the received RF signal. This bias can be significantly reduced by using high-quality components.

[0057] The radio receiver 201 may include an IQ demodulator 202 (also referred to as a quadrature demodulator) responsible for converting the RF input signal into two baseband signals, namely, an in-phase signal (component) I(t) and a quadrature signal (component) Q(t) converted into digital data by sampling. Other components of the radio receiver 201 are well known to those skilled in the art and will not be described in further detail in this disclosure.

[0058] Storage unit 205 stores one or more models adapted (in other words: adapted or trained) on RF signals originating from one or more first RF devices (e.g., first RF device 10). Storage unit 205 stores the model M adapted on the RF signals from the first RF device 10.

[0059] CPU 203 controls the operation of the third RF device 20. RAM 204 is the working memory of CPU 203.

[0060] The training (adaptation) component 206 is a software component running on the CPU 203 to adapt or train a model (e.g., M1) stored in the storage unit 205 on an RF input signal originating from a corresponding RF transmitter device (e.g., the first RF device 10).

[0061] Optionally, the third RF device 20 has a processing component 209 for preprocessing the feature data (in other words, I-data and Q-data, or alternatively, only I-data or Q-data) before providing the feature data to the training component 206. It may include a normalization component 207 and / or a filtering component 208 having one or more filters. The normalization component 207 is similar to the normalization component 307. The filtering component 208 is similar to the filtering component 308.

[0062] Figure 4 illustrates the initial phase or training procedure of the computer implementation method for authenticating the first RF device 10 as described below. The initial phase allows training of model M on RF signals transmitted from the first RF device 10. In the embodiment described below, the initial phase is performed by the second RF device 30.

[0063] In the described embodiment, model M is a machine learning (ML) model. It is also an anomaly detection (AD) model.

[0064] The training phase can be performed when both the first RF device 10 and the second RF device 30 are in a secure state, meaning that both devices 10 and 30 are trustworthy and their security is not compromised. Advantageously, the training phase can be performed at any point before the system enters the operational phase. For example, the training phase can be performed during the production of the first RF device 10, during the production of products including the RF device 10, or at a distributor or retailer (e.g., during the first field registration of the first RF device 10).

[0065] In initialization step S0, communication is established between the first RF device 10 and the second RF device 30. Each device 10 and 30 carries an ID (identifier). In step S0, the identifiers of the two devices 10 and 30 are shared through a logical agreement between the two devices 10 and 30.

[0066] In step S1, RF device 10 transmits a plurality of RF signals (e.g., a plurality of messages) to second RF device 30 via direct RF communication between the two devices 10 and 30. The distance between the two devices 10 and 30 should be appropriate.

[0067] In step S2, the second RF device 30 receives a plurality of RF signals from the first RF device 10. The RF signals are received via antenna 300 and radio receiver 301.

[0068] In step S3, the IQ demodulator 302 converts each of the received RF signals from the RF device 10 into in-phase signals I(t) and quadrature signals Q(t), and then converts them into data by sampling. A plurality of characteristic data of the received RF signals (in other words: I data and Q data, or, alternatively, only I data or Q data) are stored in memory.

[0069] In optional step S4, the feature data is filtered by the filtering unit 308 by removing and / or processing the data.

[0070] In another optional step S5, the feature data is normalized by the normalization component 307.

[0071] Steps S4 and S5 allow for the preprocessing of feature data before training step S6.

[0072] In training step S6, model M is trained on feature data (in other words, I data and Q data, or alternatively, only I data or Q data) derived from each of the plurality of RF signals received from RF device 10. The feature data may optionally be preprocessed in steps S4 and / or S5. For example, model M is trained by a second RF device 30. In this embodiment, the training is semi-supervised: it uses input training data that exclusively includes feature data obtained from RF signals transmitted from the first RF device 10. In other words, the RF signals used for training are only "normal" RF signals (in other words, RF signals from the genuine RF device 10). Therefore, model M is trained to represent normal behavior. In the context of anomaly detection, because the second RF device 30 uses RF signals from the genuine RF device 10 and the second RF device 30 predicts whether a newly received RF signal comes from the genuine device 10 or a counterfeit device, the semi-supervised training is well-adjusted.

[0073] In this variant, training is supervised. In this case, training data is labeled. Feature data (in other words: I-data and Q-data, or, alternatively, only I-data or Q-data) are labeled as "normal" or "abnormal". This assumes that both "normal" and "abnormal" RF signals are acquired and pre-labeled by the operator.

[0074] In another variant, training may be unsupervised. In this case, the feature data is unlabeled, and the second RF device 30 does not know which RF signal is abnormal, but assumes that most of the received RF signals are normal.

[0075] After training (step S6), in step S7, the trained model with reference M1 is stored in the memory 305 of the second RF device 30.

[0076] In another embodiment, the initial phase is performed by the third RF device 20. In this case, steps S2 to S6 are performed by the second RF device 30. Then, the trained model M1 is provided (e.g., transmitted) by the third RF device 20 to the second RF device 30. The second RF device 30 receives the trained model M1 and stores it in memory 305.

[0077] After the training phase, the operating mode is implemented by the second RF device 30. In the operating mode, the second RF device 30 uses the stored model M1 to determine whether the RF signal originates from the first RF device 10 from the received RF signal.

[0078] Figure 5 illustrates a flowchart of the operational phases of a computer-implemented method for authenticating an RF transmitter device 10 according to one embodiment. In one embodiment, during the operational phase, RF signals received by a second RF device 30 from the RF device T to be authenticated are converted into feature data (in other words: I data and Q data, or, alternatively, only I data or Q data). The feature data is then provided (transmitted) as input (optionally after preprocessing) to a stored (and adapted) model M1, which provides information as output regarding whether the RF device T is known (here, if it is RF device 10) or unknown.

[0079] In initialization step S00, communication is established between the RF device T to be authenticated and the second RF device 30. Each of the two devices T and 30 carries an ID (identifier) ​​shared through a logical agreement between the two devices T and 30. Therefore, the second RF device 30 receives the ID of the device T to be authenticated.

[0080] In step S10, the RF device T transmits the RF signal SG(t) to the second RF device 30 via established RF communication.

[0081] In step S11, the second RF device 30 receives the RF signal SG(t) from the RF device T via the antenna 301 and the radio receiver 302.

[0082] In one embodiment, in step S12, the IQ demodulator 302 converts the RF signal SG(t) into I data and Q data. The IQ demodulator can convert the RF signal SG(t) into in-phase signal I(t) and quadrature signal Q(t) over time, and then convert the signals I(t) and Q(t) back into I data and Q data by sampling. The feature data is directly derived from the fingerprint of the received signal SG(t) used as the RF device T. Alternatively, the RF signal SG(t) can be converted into only I data or only Q data.

[0083] In optional step S13, the I and Q data obtained from the received RF signal SG(t) are processed by one or more filters 308. More precisely, the filter unit 308 may, for example, determine whether the feature data is abnormal due to the very short duration of the signal SG(t). Abnormal feature data may be processed or ignored (to reduce or remove abnormalities). In the second case, in step S14, the second RF device 30 may request a new RF signal SG(t) from the RF device T from the RF device T, for example, after the request from the second RF device 30, and steps S10 to S13 may be repeated.

[0084] When an abnormal RF signal is detected, step S14, which requests a new RF signal from RF device T, can also be performed during the training process.

[0085] In optional step S15, as previously described, the I and Q data obtained from the received RF signal SG(t) are normalized by normalization unit 207 (preferably in cases where the feature data is not considered abnormal).

[0086] In step S16, the I-data and Q-data, optionally preprocessed in steps S13 and / or S15, are provided as input to the stored model M1. In step S16, model M1 is used in the inference phase. The model M1 stored in the second RF device 30 performs prediction based on the received I-data and Q-data as input and produces an output that is either true or false. In one embodiment, since the model is an anomaly detection model, the prediction is related to anomaly detection.

[0087] In step S17, it is determined whether the RF device T is a known device (in other words, the RF device whose trained model M1 is stored in memory 305). When the output is true, it means that the RF device T is known (here, T is the first RF device 10). When the output is false, it means that the RF device T is unknown (here, T is not the first RF device 10).

[0088] In step S18, when RF device T is known (true), the second RF device 30 allows RF device T to proceed further. For example, the second RF device 30 grants access to the application to RF device T. When RF device T is unknown (e.g., because it is a clone device), in step S19, the second RF device 30 does not allow RF device T to proceed further and can ignore any further RF signals from RF device T.

[0089] In this variant, when the RF device T is unknown (and may be a counterfeit device from an attacker), it is still allowed to proceed further but triggers an alert to the connected backend. This is intended to convince the attacker that everything is going smoothly while triggering a silent alert. Since an attacker cannot force authentication using multiple devices until one device ultimately proves useful, this feature compensates for the shortcomings of perfect detection.

[0090] In one embodiment, during the operation phase, under the control of the training component 306 running on the CPU 303, feature data derived from the RF signals received from the first RF device 10 (in other words, I data and Q data, or, alternatively, only I data or Q data) can be used to further train the corresponding stored model M1. This allows the stored model M1 to be updated with new feature data, and thus helps the model adapt to various environmental conditions not included in the initially trained model M1.

[0091] In one embodiment, model M is trained by a second RF device 30 (or a third RF device 20) on RF signals from a plurality of RF devices 10_1, 10_2, 10_3, ... . Therefore, a plurality of trained models M1, M2, M3, ... can be established by the third RF device 20 (or by the second RF device 30) using RF signals from a plurality of first RF devices 10_1, 10_2, 10_3, ... respectively. The trained models M1, M2, M3 (each trained model Mi corresponds to one RF device 10_i) are then stored in the second RF device 30, allowing identification of the corresponding RF devices 10_1, 10_2, 10_3, ... Each trained model can be stored in association with the ID (identifier) ​​of the corresponding RF device. When communication is established between the RF device T to be authenticated and the second RF device 30, the second RF device 30 receives the ID of the RF device T and can select the corresponding model Mi based on the RF device ID. Next, the selected model is used to authenticate the RF device T. Alternatively, the second RF device 30 can use the incoming RF signal from the RF device T to be authenticated to continuously try all the stored models M1, M2, M3, ... and determine whether one of them results in a true output.

[0092] In one embodiment, this authentication method is used by two RF devices to achieve mutual authentication. In this case, each of the two RF devices performs the previously defined steps S2 to S7 and S11 to S19 to authenticate the other RF device using RF signals transmitted by the other RF device. Alternatively, the training steps S2 to S6 are not performed by the RF devices themselves but by a third device, and the adapted (trained) model is then stored in the RF device.

[0093] It is possible to use only I-data or Q-data instead of using both I-data and Q-data to train model M and / or as input to the trained model during the inference phase. This allows for a reduction in the amount of data processed by the model.

[0094] In one embodiment, when two production devices have similar analog biases, they are "tagged" by production time. The two devices are then distributed to users in a way that makes it less likely for them to collide in the field (e.g., at different locations).

[0095] This disclosure also relates to computer programs containing instructions that, when executed by a computer, cause the computer to perform the steps of the previously described computer implementation method.

[0096] 10: First device; First RF device; RF device; Genuine RF device; RF transmitter device; Device; Genuine device 10_1: RF device 10_2: RF device 10_3: RF device 10_i:RF device 20: Third RF device; Third device; Device 30: Second RF device; Second device; Device 100: System 200: Antenna 201: Radio Receiver 202:IQ Demodulator 203: Processor or CPU; CPU 204:RAM 205: Storage Unit 206: Adaptor or training component; Training (adaptor) component; Training component 207: Standardized Components 208: Filtering Component 209: Processing Components 300: Antenna 301: Radio Receiver 302: IQ Demodulator 303: Processor or CPU; CPU 304: RAM 305: Storage unit; memory 306: Adaptor or training component; training (adaptor) component; training component 307: Standardized Components 308: Filtering component; filter 309: Processing Components 310: Authentication Component I(t): In-phase component; In-phase signal; Signal M: Model M1: Model; Trained Model; Adapted Model; Stored Model M2: trained model; adapted model; model; stored model M3: Trained model; Adapted model; Stored model Mi: Trained Model Q(t): orthogonal component; orthogonal signal; signal S0: Initialization steps; steps S00: Initialization steps; steps S1: Steps S2: Steps; Training steps; Receiving S3: Steps; Training steps; Results S4: Optional steps; steps; training steps S5: Optional steps; steps; training steps S6: Training steps; adaptation; steps S7: Steps; Storage S10: Steps S11: Steps; Receiving S12: Steps; Obtaining S13: Optional step; Step S14: Steps S15: Optional step; Step S16: Steps S17: Steps S18: Steps S19: Steps SG1: RF signal SG2: RF signal SG(t): New RF signal; signal; RF signal T: RF device; device to be certified; device

Claims

1. A computer implementation method for authentication of a first device (10), the computer implementation method comprising the following steps: A) receiving (S2) a plurality of RF signals (SG1, SG2, …) from the first device (10); B) obtaining (S3) feature data from the received RF signals (SG1, SG2, …); C) adapting (S6) a model (M) for authenticating the first device (10) using the obtained feature data; D) storing (S7) the adapted model (M1) in memory of a second device (30); E) receiving (S11) a new RF signal (SG(t)) from a device (T) to be authenticated via the second device (30); F) obtaining (S12) feature data from the new RF signal (SG(t)) via the second device (30); and G) The second device (30) uses the feature data obtained from the new RF signal (SG(t)) as input to the adapted model (M1) to determine whether the device to be authenticated (T) is the first device (10); wherein the feature data obtained from an RF signal includes the value of only one of the co-phase component I(t) of the RF signal over time and the quadrature component Q(t) of the RF signal over time.

2. The method of claim 1, wherein the model (M) is a machine learning model, and the step of adapting the model includes a step (S6) of using the feature data as input training data to train the machine learning model (M).

3. The method of claim 2, wherein the machine learning model (M) is trained in a semi-supervised manner using input training data that exclusively includes feature data obtained from RF signals transmitted from the first device (10).

4. The method of any one of requests 1 to 3, wherein the model (M1) is an anomaly detection model.

5. The method described in any of requests 1 to 3, wherein, After the adapted model (M1) is stored in memory, the model (M1) is updated by the second device (30) based on feature data obtained from a new RF signal (SG(t)) received from the first device (10).

6. The method of any one of claims 1 to 3 further includes a step (S5; S15) of first normalizing the feature data in at least one of steps C) and G) and then providing the normalized feature data to the model (M, M1).

7. The method of any one of claims 1 to 3 further includes a step (S4; S13) of first filtering the feature data in at least one of steps C) and G) and then providing the filtered feature data to the model (M1).

8. The method of any of claims 1 to 3, wherein steps A) to C) are performed by the second device or by a third device different from the second device.

9. The method of any one of claims 1 to 3, wherein steps A) to C) are also performed on a plurality of RF signals transmitted by the second device (30) to adapt to a model (M2) for authenticating the second device (30), and steps D) to G) are also performed by the first device (10) for authenticating the second device (30) using the adapted model (M2), such that the first device and the second device perform a mutual authentication.

10. The method of any one of claims 1 to 3, wherein steps A) to C) are performed on a plurality of first devices, and a plurality of adapted models (M1, M2, M3, …) corresponding to the plurality of first devices are stored in memory in the second device (30).

11. The method of claim 10, further comprising a step of establishing a communication between the device to be authenticated (T) and the second device (30), wherein the device to be authenticated (T) transmits an identifier to the second device; and a step of the second device (30) selecting, based on the identifier of the device to be authenticated, an adapted model corresponding to the device to be authenticated (T) from among the plurality of adapted models stored in memory, the selected adapted model being used in step G).

12. A system for authenticating a first device (10) comprising a second device (30) configured to perform steps A) to G) of the method defined in Request 1, or a second device (30) configured to perform steps D) to G) of the method defined in Request 1, and a third device (20) configured to perform steps A) to C) of the method defined in Request 1.

13. A computer program containing instructions that, when executed by a computer, cause the computer to perform the steps defined in request 1.

Images