A distributed secure computing method based on cloud-edge service platform

Abstract

The present invention proposes a distributed secure computing method based on a cloud-edge service platform, which adopts a three-layer network structure of cloud center server-edge node-user terminal device. The method includes: the cloud center server broadcasts a global model to multiple user terminal devices via multiple edge nodes; each user terminal device calculates the corresponding local gradient using its own data set for the global model, thereby obtaining a local gradient standard value, and uses a maximum power transmission method to send the local gradient standard value to the associated edge node; each edge node obtains a received signal based on the local gradient standard value and transmission power of the associated user terminal device, obtains its own edge node gradient estimation value based on the received signal and sends it to the cloud center server; the cloud center server aggregates the gradient estimation values ​​of each edge node to update the global model to obtain the optimal global model. The method of the present invention can reduce the impact of Byzantine attacks and enhance the robustness of transmission from the user terminal to the edge node.

Description

Technical Field

[0001] The present invention relates to the field of cloud computing, and in particular to a distributed secure computing method based on a cloud-edge service platform. Background Art

[0002] In the field of edge cloud computing, nodes in the network can share computing power and data. Using the data of each node to train machine learning models, useful information can be mined from the data, thereby supporting applications such as autonomous driving and smart cities. Among them, federated learning for edge computing can utilize rich distributed data while protecting user privacy through a distributed computing method that only transmits local models and does not transmit local raw data. However, when there are too many network nodes participating in distributed computing and the parameter dimension of the learning model is too high, communication between nodes will become a bottleneck. In order to solve the communication bottleneck problem, federated learning over the air (FLOA) based on simulated aggregation has become a promising technology that provides higher communication efficiency and privacy protection. This method involves concurrent transmission and aggregation, in which edge network devices (workers) upload their local updates to the parameter server (PS) simultaneously through shared time-frequency resources. Then, the PS only obtains the average update instead of a single local update, thereby reducing latency and communication costs. However, this simultaneous process can put FLOA into security risks, especially Byzantine attacks, where individual malicious nodes upload incorrect models to disrupt the learning process.

[0003] In addition, most existing federated learning adopts a two-layer architecture, namely the user end and the server (cloud server or edge server). Moreover, the number of participating devices in federated learning based on simulated aggregate communication is often small. The reason is that simulated aggregate communication requires high synchronization, that is, distributed user-end devices need to send their local models at the same time. However, for devices that are far away, synchronization encounters challenges. Most existing technologies are aimed at federated learning under digital communication, and use statistical principles to build a screening and removal mechanism to exclude Byzantine attackers, such as mean, median, etc. Those far from the median are judged as malicious attackers. The disadvantage is that the central node must obtain the model parameter value sent by each user in order to know which node sends the parameter far from the average. However, federated learning based on simulated aggregate communication can only obtain aggregate values ​​at the center, not individual values. This makes the current Byzantine attack defense measures ineffective. Summary of the invention

[0004] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0005] To this end, the first objective of the present invention is to propose a distributed secure computing method based on a cloud-edge service platform to reduce the impact of Byzantine attacks and enhance the robustness of transmission from the user end to the edge node.

[0006] The second objective of the present invention is to propose a cloud-edge service platform.

[0007] The third object of the present invention is to propose a distributed secure computing device based on a cloud-edge service platform.

[0008] A fourth object of the present invention is to provide a computer-readable storage medium.

[0009] To achieve the above-mentioned purpose, the first aspect of the present invention proposes a distributed secure computing method based on a cloud-edge-end service platform, which adopts a three-layer network structure of a cloud center server-edge node-user terminal device. The method includes:

[0010] The cloud center server broadcasts the global model to multiple user-end devices via multiple edge nodes;

[0011] Each user terminal device calculates the corresponding local gradient using its own data set for the global model, thereby obtaining a local gradient standard value, and sends the local gradient standard value to the associated edge node using a maximum power transmission method;

[0012] Each edge node obtains a received signal based on the local gradient standard value and transmission power of the associated user terminal device, obtains a respective edge node gradient estimation value based on the received signal and sends the estimated value to the cloud center server;

[0013] The cloud center server aggregates the estimated gradient values ​​of each edge node to update the global model, thereby obtaining the optimal global model.

[0014] In the method of the first aspect of the present invention, each user-end device calculates the corresponding local gradient using its own data set for the global model, and then obtains the local gradient standard value, including: after each user-end device receives the global model, it uses its own data set for training to calculate the corresponding local gradient, and sends the relevant value of the local gradient to the associated edge node; each edge node uses a simulated aggregation transmission model to obtain a global estimation value based on the relevant value of the local gradient of the associated user-end device, and returns the global estimation value to all user-end devices; each user-end device standardizes the local gradient based on the global estimation value to obtain the local gradient standard value.

[0015] In the method of the first aspect of the present invention, each edge node adopts a simulated aggregation transmission model to obtain a global estimation value based on the correlation value of the local gradient of the associated user terminal device, including: each edge node calculates the global mean and global variance of the gradient for the correlation value of the local gradient of all associated user terminal devices, and the global mean and global variance constitute the global estimation value.

[0016] In the method of the first aspect of the present invention, when the user terminal device transmits the local gradient standard value using the maximum power transmission mode, the transmission power of each user terminal device is determined based on the maximum power and the dimension of the model parameters.

[0017] In the method of the first aspect of the present invention, the cloud center server uses a geometric median aggregation method to aggregate the gradient estimation values ​​of each edge node to update the global model.

[0018] To achieve the above-mentioned purpose, the second aspect of the present invention proposes a cloud-edge service platform, which includes a cloud center server, multiple edge nodes and multiple user-end devices;

[0019] Each user terminal device is used to calculate the corresponding local gradient using the respective data set for the received global model, thereby obtaining a local gradient standard value, and sending the local gradient standard value to the associated edge node using a maximum power transmission method;

[0020] Each edge node is used to obtain a received signal based on a local gradient standard value and a transmission power of an associated user terminal device, obtain a respective edge node gradient estimation value based on the received signal, and send the estimated value to a cloud center server;

[0021] The cloud center server is used to broadcast the global model to each user terminal device via the edge node; it is also used to aggregate the gradient estimation values ​​of each edge node to update the global model, so as to obtain the optimal global model.

[0022] In the cloud-edge service platform of the second aspect of the present invention, in each edge node, the maximum power transmission method refers to determining the transmission power based on the dimensions of maximum power and model parameters, and then sending the local gradient standard value to the associated edge node at the transmission power.

[0023] In the cloud-edge service platform of the second aspect of the present invention, the aggregation method adopted by the cloud center server is the geometric median aggregation method.

[0024] To achieve the above-mentioned purpose, the third aspect of the embodiment of the present invention proposes a distributed secure computing device based on a cloud-edge service platform, including: a processor, and a memory communicatively connected to the processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory to implement the method proposed in the first aspect of the present invention.

[0025] To achieve the above-mentioned purpose, the fourth aspect of the present invention proposes a computer-readable storage medium, in which computer-readable storage medium is stored computer execution instructions, and when the computer execution instructions are executed by a processor, they are used to implement the method proposed in the first aspect of the present invention.

[0026] The distributed secure computing method, device, cloud edge service platform and storage medium based on the cloud edge service platform provided by the present invention adopt a three-layer network structure of cloud center server-edge node-user terminal device, and broadcast a global model to multiple user terminal devices through a cloud center server via multiple edge nodes; each user terminal device calculates the corresponding local gradient using its own data set for the global model, thereby obtaining a local gradient standard value, and uses the maximum power transmission method to send the local gradient standard value to the associated edge node; each edge node obtains a received signal based on the local gradient standard value and transmission power of the associated user terminal device, obtains its own edge node gradient estimation value based on the received signal and sends it to the cloud center server; the cloud center server aggregates the gradient estimation values ​​of each edge node to update the global model, thereby obtaining the optimal global model. In this case, a three-layer network structure of cloud center server-edge node-user terminal device is adopted, and the user terminal device uses the maximum power transmission method to send the calculated local gradient standard value to the associated edge node, the edge node performs air calculation to obtain the edge node gradient estimation value, and the cloud center server aggregates the gradient estimation values ​​of each edge node to update the global model. As a result, the impact of Byzantine attacks is reduced and the robustness of transmission from the user terminal to the edge node is enhanced.

[0027] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0029] Figure 1 A schematic diagram of a cloud-edge-end three-layer network architecture provided in an embodiment of the present invention;

[0030] Figure 2 A schematic diagram of a process flow of a distributed secure computing method based on a cloud-edge-end service platform provided in an embodiment of the present invention;

[0031] Figure 3 A block diagram of the cloud-edge service platform provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0033] The following describes a distributed secure computing method, device, cloud-edge-end service platform and storage medium based on a cloud-edge-end service platform according to an embodiment of the present invention with reference to the accompanying drawings.

[0034] The embodiment of the present invention provides a distributed secure computing method based on a cloud-edge service platform to reduce the impact of Byzantine attacks and enhance the robustness of transmission from the user end to the edge node.

[0035] In the present invention, the distributed secure computing method based on the cloud-edge-end service platform adopts a three-layer network structure of cloud center server-edge node-user terminal device. The three-layer network structure of cloud center server-edge node-user terminal device is referred to as cloud-edge-end three-layer network architecture. The cloud-edge-end three-layer network architecture includes a cloud center server, multiple edge nodes and multiple user terminal devices.

[0036] Figure 1 Schematic diagram of the cloud-edge-end three-layer network architecture provided by an embodiment of the present invention. Figure 1 As shown, the cloud-edge-end three-layer network architecture includes a cloud center server (i.e., cloud center PS, also called cloud center parameter server), multiple edge nodes, and multiple user-end devices. Specifically, the cloud center server is connected to multiple edge nodes in communication, each edge node is connected to a group of user-end devices (i.e., each edge node is associated with a corresponding group of user-end devices), and each group of user-end devices includes at least one user-end device. User-end devices are also called user terminals, local devices, etc.

[0037] In an embodiment of the present invention, the number of edge nodes in the cloud-edge-end three-layer network architecture is represented as V, and the number of user-end devices is represented as U. U user-end devices are divided into V groups, and each group of user-end devices is connected to different edge nodes. The number of user-end devices in each group of user-end devices is multiple. For example, the number of user-end devices connected to the vth edge node is Uv. v ranges from 1 to V. Each user-end device has a local database, and the number of data points contained in the local database is K. These data points of different user-end devices are from the same large data set. The kth data point of the i-th user terminal device can be expressed as (x i,k ,y i,k ).

[0038] Figure 2 A flow chart of a distributed secure computing method based on a cloud-edge service platform provided in an embodiment of the present invention.

[0039] like Figure 2 As shown, the distributed secure computing method based on the cloud-edge service platform includes the following steps:

[0040] Step S101: the cloud center server broadcasts the global model to multiple user-end devices via multiple edge nodes.

[0041] Specifically, in step S101, when federated learning is required, the cloud center server broadcasts the initial global model to each edge node; each edge node broadcasts the received initial global model to the associated user terminal device.

[0042] Here we introduce federated learning in combination with the cloud-edge-end three-layer network architecture:

[0043] In the cloud-edge-end three-layer network architecture, the method of the present invention considers a network consisting of a cloud center PS, V edge nodes and U user-end devices. Each user-end device has a local database containing K data points, which are collected from a large data set. The symbol (x i,k ,y i,k ) represents the kth data point of the i-th user terminal device. Introducing the loss function f(w; x i,k ,y i,k ) to quantify the error associated with each data point, where the model parameter vector satisfies w = [w 1 ,…,w D ], the model parameter vector includes model parameters of dimension D. Through collaboration, the cloud center PS, edge nodes and user-end devices aim to learn the model parameter vector w by minimizing the expected value of the loss function, which is in the dataset The above is expressed as the overall loss function F(w), which is the sum of all loss functions f(w; x i,k ,y i,k ), the mean (i.e. expected value) of the loss function f(w; x i,k ,y i,k ) is the loss of the kth data point of the i-th user terminal device. The overall loss function F(w) satisfies:

[0044] w * =argmin w F(w) (1)

[0045] In the formula, w * is the optimal model parameter vector, argmin w F(w) represents the model parameter vector that minimizes the overall loss function F(w).

[0046] Typically, the stochastic gradient descent (SGD) algorithm is used to minimize F(w) in a two-layer network architecture to achieve federated learning (FL). In this way, the model parameter vector w at the tth iteration is tThe update formula at the cloud center PS is:

[0047]

[0048] In the formula, α is the learning rate, w t-1 is the model parameter vector at the t-1th iteration, represents the local gradient calculated by the i-th client device at the t-th iteration by using a randomly selected data sample (e.g., the k-th data point). Note that the local gradient is based on the model parameter vector w from the previous iteration (t-1). t-1 In order for the cloud center PS to obtain the sum of these local gradients, a good communication and aggregation scheme must be established. This is crucial for effective collaboration and model parameter updates in the federated learning framework.

[0049] Step S101 is the triggering step of federated learning, that is, when the cloud center server receives the instruction to perform federated learning, it broadcasts the initial global model to multiple user end devices through multiple edge nodes. The parameters of the global model are the model parameter vector w that needs to be optimized. The processing steps such as local gradient and aggregation scheme are described in detail later.

[0050] In step S102, each user terminal device calculates the corresponding local gradient using its own data set for the global model, thereby obtaining a local gradient standard value, and sends the local gradient standard value to the associated edge node using a maximum power transmission method.

[0051] In step S102, each user terminal device calculates the corresponding local gradient using its own data set for the global model, and then obtains the local gradient standard value, including: after receiving the global model, each user terminal device uses its own data set (i.e., local database) for training to calculate the corresponding local gradient, and sends the relevant value of the local gradient to the associated edge node; each edge node uses the simulated aggregation transmission model to obtain a global estimated value based on the relevant value of the local gradient of the associated user terminal device, and returns the global estimated value to all user terminal devices; each user terminal device performs standard processing on the local gradient based on the global estimated value to obtain the local gradient standard value. Among them, each edge node uses the simulated aggregation transmission model to obtain the global estimated value based on the relevant value of the local gradient of the associated user terminal device, including: each edge node calculates the global average value and global variance of the gradient for the relevant value of the local gradient of all associated user terminal devices, and the global average value and the global variance constitute the global estimated value.

[0052] In step S102, when the user terminal device transmits the local gradient standard value in the maximum power transmission mode, the transmission power of each user terminal device is determined based on the maximum power and the dimension of the model parameter.

[0053] Specifically, in order to reflect the potential changes in gradient statistics during iteration, normalization is performed in each round of communication (i.e., each iteration). More precisely, in each round of communication, after receiving the global model, each user terminal device uses its own data set (i.e., local database) for training to calculate the corresponding local gradient and the related value of the local gradient. Taking the i-th user terminal device at the t-th iteration as an example, the local gradient g of the i-th user terminal device at the t-th iteration is i,t The relevant values ​​include the gradient mean and gradient variance. The gradient mean of the i-th user terminal device at the t-th iteration is satisfy: D is the model parameter dimension, is the d-th dimension parameter of the local gradient update of the i-th user-end device at the t-th iteration, and the gradient variance of the i-th user-end device at the t-th iteration satisfy: The relevant values ​​of local gradients are key parameters in the subsequent process.

[0054] Then, each user-end device shares the relevant values ​​of its own local gradient (i.e., the gradient mean and the gradient variance) with the corresponding edge node. The edge node uses air calculation to perform effective gradient aggregation, and the relevant values ​​of the local gradient of the user-end device are analog transmitted through amplitude modulation. These modulated signals are transmitted simultaneously from the user-end device to the edge node using the same multiple access channel. The present invention assumes that symbol-level synchronization is achieved between local devices through a dedicated synchronization channel. For symbol-level synchronization, this is because multiple user-end devices are transmitting simultaneously and synchronization is required. Information is transmitted in a bit stream, and the present invention transmits one-dimensional model parameters according to one symbol for each bit. Symbol-level synchronization means that each parameter transmission is synchronized between multiple user-end devices, that is, it is assumed that a synchronization method can achieve simultaneous transmission of multiple devices and each symbol is aligned.

[0055] After receiving the relevant values ​​of these local gradients, each edge node aggregates them by calculating the global mean and global variance of the gradients using the simulated aggregation transmission model. Taking the vth edge node as an example, the global mean calculated by the vth edge node satisfies: In the formula, is the gradient average of the i-th user-end device associated with the v-th edge node at the t-th iteration, and Uv is the number of user-end devices connected to the v-th edge node. is the global average value of the vth edge node at the tth iteration. The global variance calculated by the vth edge node satisfies: In the formula, is the gradient variance of the i-th user terminal device associated with the v-th edge node at the t-th iteration. is the global variance of the vth edge node at the tth iteration. Subsequently, each edge node broadcasts the calculated global estimates (i.e., global mean and global variance) back to all local devices. These global estimates are then used in the normalization process.

[0056] After receiving the global estimated value (i.e., the global mean and global variance), each user terminal device performs standardization on the local gradients of each user terminal device at different iterations. Taking the i-th user terminal device at the t-th iteration as an example, the local gradient standard value satisfies:

[0057]

[0058] In the formula, is the local gradient standard value of the i-th user terminal device at the t-th iteration, g i,t is the local gradient of the i-th user-end device at the t-th iteration, represents a vector consisting entirely of 1s, with the same dimension as g i,t Same, is the global average value received by the i-th user terminal device at the t-th iteration, ∈ t is the square root of the global variance received by the i-th user terminal device at the t-th iteration. represents a D-dimensional vector, in, is the standard value of the d-th dimension parameter of the local gradient update of the i-th user terminal device at the t-th iteration. In order to optimize the subsequent power control strategy, the standardized It can ensure zero mean and unit variance, that is, This normalization enables edge nodes to design power control strategies without the need to know specific transmitted symbols. Notably, the normalization factor remains consistent across all local gradients, allowing edge nodes to perform inversion.

[0059] In step S102, in each round of communication, the focus is on the global estimate (i.e., the global mean and global variance), and the gradient mean and gradient variance of each user end device are collected in turn on the edge node. This process ensures that one individual contribution to the global statistics is collected at a time. It should be noted that these communications specifically used for standardization are considered to be noise-free in the present invention, which means that they do not introduce errors when they occur. Because the amount of data is small, error control is easier to achieve.

[0060] In step S102, considering the presence of Byzantine attackers on the user-side devices, the Byzantine attackers have knowledge about the prescribed normalization method. In order to hide their maliciousness during the normalization process, these attackers choose to transmit the exact mean and variance of their local gradients. This move is intended to avoid detection, as any deviation from the expected behavior may expose them. The rationale behind this strategy is to prevent edge nodes from identifying and subsequently filtering Byzantine attackers. This is especially important because both normal workers and Byzantine workers operate on independent and identically distributed (iid) datasets, which makes it crucial for attackers to mimic real behavior to evade detection.

[0061] In general, all user-end devices standardize their local gradients (i.e., local gradient standard values) ) is transmitted to an edge node with a certain transmission power. The transmission of each user terminal device is subject to the transmission power constraint. Taking the i-th user terminal device as an example, the i-th user terminal device obtains the local gradient standard value When transmitting to the corresponding edge node, the transmission uses the p i,t It is performed at a specific transmit power level represented by . Note that there is a constraint on the transmission power of the ith user terminal device. This constraint regulates the amount of power used during the transmission, ensuring that it meets the following predefined constraints:

[0062]

[0063] In the formula, p i,t is the transmission power of the i-th user terminal device at the t-th iteration, E[] represents the expectation, is the maximum power of the ith user-end device. Simplifying equation (4), the power limit of the ith user-end device at the tth iteration can be expressed as This limit is essential to regulating the transmit power of user-side equipment, ensuring it remains within specified limits for efficient and controlled communications.

[0064] For a Byzantine attacker, when updating the gradient to the edge node Provides an arbitrary value, which may interfere with FL. The transmit power controlled by the nth Byzantine attacker obeys the constraint:

[0065]

[0066] In the formula, is the transmission power of the nth Byzantine attacker at the tth iteration, is the local gradient of the nth Byzantine attacker at the tth iteration, is the maximum power of the nth Byzantine attacker.

[0067] In order to deal with potential intelligent Byzantine attacks, the method of the present invention optimizes the predefined protocols used between the cloud center PS, edge nodes and local devices. The optimized predefined protocol stipulates the strategy of maximum power transmission mode to actively mitigate the impact through normal user-end devices to ensure that federated learning develops in the right direction. This strategy, called maximum power transmission mode, requires normal user-end devices to transmit their local gradients at maximum transmission power regardless of their channel status. This initiative aims to enhance the robustness of federated learning against intelligent Byzantine attackers (who may also transmit false local gradients at maximum power) by leveraging the collective efforts of normal devices and their transmission of local updates at maximum power. In the maximum power transmission mode, the transmission power of the i-th local device associated with the v-th edge node at the t-th iteration is expressed as:

[0068]

[0069] In the formula, p v,i,t is the transmit power (transmission power) of the ith local device associated with the vth edge node at the tth iteration, is the maximum power of the ith local device associated with the vth edge node.

[0070] Step S103, each edge node obtains a received signal based on the local gradient standard value and transmission power of the associated user terminal device, obtains its own edge node gradient estimation value based on the received signal and sends it to the cloud center server.

[0071] In step S103, considering the case where there are Byzantine attackers in the user terminal device, taking all user terminal devices (i.e., U user terminal devices) as an example, there are N Byzantine attackers and M (M=UN) normal user terminal devices, and considering the scenario with block fading channels in the wireless FL, where the wireless channel remains unchanged in each iteration, but may change independently between iterations. Each iteration represented by index t corresponds to a time block. During the tth iteration, the received signal obtained by simulating aggregate transmission at the vth edge node is expressed as:

[0072]

[0073] where y and y correspond to contributions from normal operators (i.e., normal user-end devices), attackers, and noise, respectively. Specifically, y v,t represents the received signal of the vth edge node at the tth iteration, M v is the number of normal user-end devices among all user-end devices associated with the vth edge node, p v,m,tis the transmission power of the mth normal user terminal device among all normal user terminal devices associated with the vth edge node at the tth iteration, |h v,m,t | is the channel gain from the mth normal user terminal device to the vth edge node at the tth iteration, is the local gradient standard value of the mth normal user terminal device among all normal user terminal devices associated with the vth edge node at the tth iteration. v is the number of Byzantine attackers in all client devices associated with the vth edge node, |h v,n,t | is the channel gain from the nth Byzantine attacker to the vth edge node at the tth iteration, is the local gradient of the nth Byzantine attacker among all Byzantine attackers associated with the vth edge node at the tth iteration, is the transmission power of the nth Byzantine attacker among all Byzantine attackers associated with the vth edge node at the tth iteration, z v,t is an additive white Gaussian noise (AWGN) that is independent of the gradient update. The channel exhibits independent Rayleigh fading. The channel is assumed to be fully known at the user end device and the edge node, allowing the channel phase offset to be compensated by the user end device before transmitting the gradient update.

[0074] Taking the vth edge node as an example, the vth edge node denormalizes the received signal to obtain the estimated total gradient (i.e., the vth edge node gradient estimate ). This process involves reversing the normalization applied in the normalized local update (3). Denormalization is expressed as follows:

[0075]

[0076] Where M v +N v =U v , U v is the total number of local devices associated with the vth edge node, where The present invention allows all edge nodes to collect gradient updates from the same number of local devices. The terms on the right side of the above equation represent different contributions: the first term represents the aggregated gradient from normal local devices, the second and third terms represent the malicious contribution of the Byzantine attacker to the gradient update, and the last term is Gaussian white noise. v,t is the square root of the global variance calculated for the vth edge node, p v,i,t is the transmission power of the i-th user terminal device associated with the v-th edge node at the t-th iteration, |h v,i,t | is the channel gain from the ith user end device to the vth edge node at the tth iteration, and the channel exhibits independent Rayleigh fading. is the global average value calculated by the vth edge node at the tth iteration, g v,m,t is the local gradient of the mth normal user terminal device among all normal user terminal devices associated with the vth edge node at the tth iteration.

[0077] In this embodiment, when step S103 executes equation (8), based on the maximum power transmission mode, the transmission power p of the mth normal user terminal device among all normal user terminal devices associated with the vth edge node at the tth iteration is v,m,t The value of (6) is taken because it can be clearly seen from the simulated aggregate transmission that the power control strategy will affect the final received signal-to-noise ratio and the signal strength difference between normal workers and attackers, thereby affecting the learning performance. Therefore, the maximum power transmission method can actively mitigate the impact through normal local devices and ensure that federated learning develops in the right direction.

[0078] In step S103, each edge node obtains the corresponding edge node gradient estimation value using formula (8), and then sends it to the cloud center PS through a digital link for final aggregation to obtain a global model.

[0079] In step S104, the cloud center server aggregates the estimated gradient values ​​of each edge node to update the global model, thereby obtaining the optimal global model.

[0080] In step S104, the cloud center server uses the geometric median aggregation method to aggregate the gradient estimation values ​​of each edge node to update the global model.

[0081] Specifically, since only a small number of edge nodes are considered, the focus of the present invention is to simulate aggregate transmission, in which the link between the edge node and the cloud center PS is considered to be error-free (this can be achieved through error control or retransmission in digital communication). In addition, since the embodiment of the present invention adopts a cloud-edge-end three-layer network architecture, the cloud center PS receives the gradient estimate of the edge node, so it is not suitable to use formula (2) for model update. In an embodiment of the present invention, the cloud center PS can use many effective aggregation methods to aggregate the gradient estimates of each edge node, such as geometric median, coordinate median, coordinate trimmed mean, etc. In the embodiment of the present invention, the geometric median is used, so that the global model is updated by formula (9):

[0082]

[0083] Where α is the learning rate, and "geomed" (.) represents the geometric median aggregation, which is defined as:

[0084]

[0085] In the formula, is the estimated gradient calculated for the vth edge node, and g is the global model gradient variable.

[0086] The embodiment of the present invention utilizes geometric median aggregation as an effective countermeasure against Byzantine attacks, especially when a single local update is transmitted separately to the cloud center PS. When the number of Byzantine attackers in the user terminal device is less than half of the total number, the method of the present invention is very effective in approximating the average value of the true gradient update. In addition, in the present invention, if there are Byzantine edge nodes in the edge nodes, the method of the present invention can also be extended to the update of the aggregation group (i.e., edge node), provided that the number of Byzantine edge nodes should be less than half of the total number of edge nodes V, so as to accurately approximate the average value of the update of the aggregation group that only includes normal user terminal devices.

[0087] Considering that the Weiszfeld algorithm can improve the calculation efficiency of the geometric median, in order to solve the problem of numerical instability, the embodiment of the present invention uses a smoothed version of the Weiszfeld algorithm to calculate the geometric median, which is expressed as:

[0088]

[0089] in,

[0090]

[0091] Where ι is the smoothing factor, ι>0,

[0092] The specific steps of calculating the geometric median using the smoothed version of the Weiszfeld algorithm include: 1) Given a smoothing factor ι>0, update the value g according to the received gradient v,t , calculated for each edge node 2): Calculate the global model gradient variable based on the calculation results 3): Repeat the above two steps until the global model gradient variable g converges, thereby obtaining the optimal value of the global model gradient variable.

[0093] Substitute the optimal value of the global model gradient variable into equation (9) to update the model parameter vector to obtain the model parameter vector w at the tth iteration: t , update the number of iterations, repeat the above steps S102 to S104 to obtain the model parameter vector of the new number of iterations, until the maximum number of iterations T is reached. Based on the model parameter vector of each iteration number, the optimal model parameter vector (i.e., the optimal global model) is obtained using formula (1).

[0094] In order to implement the above method embodiments, the present invention proposes a cloud-edge service platform.

[0095] Figure 3This is a block diagram of the cloud-edge service platform provided by an embodiment of the present invention. Figure 3 As shown, the cloud-edge service platform includes a cloud center server, multiple edge nodes and multiple user-end devices, wherein:

[0096] Each user terminal device is used to calculate the corresponding local gradient using the respective data set for the received global model, thereby obtaining the local gradient standard value, and sending the local gradient standard value to the associated edge node using the maximum power transmission mode;

[0097] Each edge node is used to obtain a received signal based on a local gradient standard value and a transmission power of an associated user terminal device, obtain a respective edge node gradient estimation value based on the received signal, and send the estimated value to a cloud center server;

[0098] The cloud center server is used to broadcast the global model to each user terminal device via the edge node; it is also used to aggregate the gradient estimation values ​​of each edge node to update the global model, thereby obtaining the optimal global model.

[0099] Furthermore, in a possible implementation of an embodiment of the present invention, in each edge node, the maximum power transmission method refers to determining the transmission power based on the dimensions of the maximum power and the model parameters, and then sending the local gradient standard value to the associated edge node with the transmission power.

[0100] Furthermore, in a possible implementation of an embodiment of the present invention, the aggregation method adopted by the cloud center server is a geometric median aggregation method.

[0101] It should be noted that the aforementioned explanation of the embodiment of the distributed secure computing method based on the cloud-edge service platform also applies to the cloud-edge service platform of this embodiment and will not be repeated here.

[0102] In the embodiment of the present invention, a three-layer network structure of cloud center server-edge node-user terminal device is adopted, and a global model is broadcasted to multiple user terminal devices through a cloud center server via multiple edge nodes; each user terminal device calculates the corresponding local gradient using its own data set for the global model, thereby obtaining a local gradient standard value, and uses the maximum power transmission method to send the local gradient standard value to the associated edge node; each edge node obtains a received signal based on the local gradient standard value and transmission power of the associated user terminal device, obtains its own edge node gradient estimation value based on the received signal and sends it to the cloud center server; the cloud center server aggregates the gradient estimation values ​​of each edge node to update the global model, thereby obtaining the optimal global model. In this case, a three-layer network structure of cloud center server-edge node-user terminal device is adopted, and the user terminal device uses the maximum power transmission method to send the calculated local gradient standard value to the associated edge node, the edge node performs air calculation to obtain the edge node gradient estimation value, and the cloud center server aggregates the gradient estimation values ​​of each edge node to update the global model. As a result, the impact of Byzantine attacks is reduced and the robustness of transmission from the user terminal to the edge node is enhanced.

[0103] Since the edge nodes can be closer to the distributed devices, they provide a stable communication link for the parameter transmission of the distributed devices. As a relay, the edge nodes can allow more devices to participate in FL. Therefore, the method of the present invention introduces an innovative framework that uses a cloud-edge-end three-layer network to help realize heterogeneous FLOA. The framework mitigates the impact of Byzantine attacks while retaining the advantages of air computing to achieve efficient and secure federated learning. In the method of the present invention, the edge nodes perform air computing, collect the trained model gradients from the user-end devices and aggregate them. Subsequently, the edge nodes transmit these aggregated gradients to the cloud center PS in different time slots, and the PS applies a robust aggregation technique to combine the updates from the edge nodes. The edge nodes are used to collect the models trained by each local user-end device, which can expand the number of participating devices and solve the problem of difficult synchronization between devices at long distances. By adopting the form of edge node grouping for model aggregation, the attacked edge group can be screened out. In order to enhance the robustness of the transmission from the user-end device to the edge node, the present invention uses a maximum power transmission method to avoid malicious attacks to the greatest extent. The present invention solves the Byzantine attack problem existing in the edge-cloud computing system using simulated aggregation transmission technology, and also solves the limitations of the existing technology, promotes the development of cloud-edge federated learning technology, and provides a more flexible, efficient and secure solution.

[0104] In order to implement the above embodiments, the present invention also proposes a distributed secure computing device based on a cloud-edge-end service platform, comprising: a processor, and a memory connected to the processor in communication; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to implement the method provided in the above embodiments. The processor and the memory are both multiple and distributed in the cloud center server, each edge node and each user terminal device.

[0105] In order to implement the above embodiments, the present invention further proposes a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, they are used to implement the methods provided by the above embodiments.

[0106] In order to implement the above embodiments, the present invention further provides a computer program product, including a computer program, which implements the methods provided by the above embodiments when executed by a processor.

[0107] The collection, storage, use, processing, transmission, provision and disclosure of user personal information involved in this invention shall comply with the relevant laws and regulations and shall not violate public order and good morals.

[0108] It should be noted that personal information from users should be collected for legitimate and reasonable purposes and should not be shared or sold outside of these legitimate uses. In addition, such collection / sharing should be carried out after receiving the user's informed consent, including but not limited to notifying the user to read the user agreement / user notice and sign the agreement / authorization including authorization of relevant user information before the user uses the function. In addition, any necessary steps should be taken to protect and safeguard access to such personal information data and ensure that others who have access to personal information data comply with its privacy policy and procedures.

[0109] The present invention contemplates providing embodiments where users can selectively block the use or access of personal information data. That is, the present disclosure contemplates providing hardware and / or software to prevent or block access to such personal information data. Once the personal information data is no longer needed, risks can be minimized by limiting data collection and deleting the data. In addition, when applicable, such personal information is de-identified to protect the privacy of the user.

[0110] In the description of the aforementioned embodiments, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0111] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0112] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present invention belong.

[0113] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0114] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0115] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0116] In addition, each functional unit in each embodiment of the present invention may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0117] The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present invention. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A distributed secure computing method based on a cloud-edge-end service platform, characterized in that: Using a three-layer network structure of cloud center server-edge node-user terminal device, the method includes: The cloud center server broadcasts the global model to multiple user-end devices via multiple edge nodes; Each user terminal device calculates the corresponding local gradient using its own data set for the global model, thereby obtaining a local gradient standard value, and sends the local gradient standard value to the associated edge node using a maximum power transmission method; Each edge node obtains a received signal based on the local gradient standard value and transmission power of the associated user terminal device, obtains a respective edge node gradient estimation value based on the received signal and sends it to the cloud center server, wherein when the user terminal device transmits the local gradient standard value using the maximum power transmission mode, the transmission power of each user terminal device is determined based on the maximum power and the dimension of the model parameter; the method for determining the received signal is: in, is the received signal of the vth edge node at the tth iteration, is the number of normal user-end devices among all user-end devices associated with the vth edge node, is the transmission power of the mth normal user terminal device among all normal user terminal devices associated with the vth edge node at the tth iteration, is the channel gain from the mth normal user terminal device to the vth edge node at the tth iteration, is the local gradient standard value of the mth normal user terminal device among all normal user terminal devices associated with the vth edge node at the tth iteration, is the number of Byzantine attackers in all client devices associated with the vth edge node, At the tth iteration, The channel gain from a Byzantine attacker to the vth edge node is, is the number of Byzantine attackers associated with the vth edge node at the tth iteration. The local gradient of a Byzantine attacker, is the number of Byzantine attackers associated with the vth edge node at the tth iteration. The transmission power of a Byzantine attacker, is an additive Gaussian white noise that is independent of the gradient update; The cloud center server aggregates the estimated gradient values ​​of each edge node to update the global model, thereby obtaining the optimal global model.

2. The distributed secure computing method based on the cloud-edge-end service platform according to claim 1 is characterized in that: Each of the user end devices calculates the corresponding local gradient using its own data set for the global model, and then obtains the local gradient standard value, including: After receiving the global model, each user terminal device uses its own data set for training to calculate the corresponding local gradient, and sends the relevant value of the local gradient to the associated edge node; Each edge node uses a simulated aggregation transmission model to obtain a global estimation value based on the correlation value of the local gradient of the associated user terminal device, and returns the global estimation value to all user terminal devices; Each user terminal device performs standardization processing on the local gradient based on the global estimated value to obtain a local gradient standard value.

3. The distributed secure computing method based on the cloud-edge-end service platform according to claim 2 is characterized in that: Each edge node uses a simulated aggregation transmission model to obtain a global estimation value based on a correlation value of a local gradient of an associated user terminal device, including: Each edge node calculates a global average value and a global variance of the gradient for the correlation values ​​of the local gradients of all associated user-end devices, and the global average value and the global variance constitute the global estimation value.

4. The distributed secure computing method based on the cloud-edge-end service platform according to claim 1 is characterized in that: The cloud center server uses a geometric median aggregation method to aggregate the gradient estimation values ​​of each edge node to update the global model.

5. A cloud-edge-device service platform, characterized in that: The cloud-edge service platform includes a cloud center server, multiple edge nodes and multiple user-end devices; Each user terminal device is used to calculate the corresponding local gradient using the respective data set for the received global model, thereby obtaining a local gradient standard value, and sending the local gradient standard value to the associated edge node using a maximum power transmission method; Each edge node is used to obtain a received signal based on a local gradient standard value and a transmission power of an associated user terminal device, obtain a respective edge node gradient estimation value based on the received signal, and send the estimated value to a cloud center server; The maximum power transmission mode refers to determining the transmission power based on the dimension of the maximum power and the model parameter, and then sending the local gradient standard value to the associated edge node with the transmission power; the method for determining the received signal is: in, is the received signal of the vth edge node at the tth iteration, is the number of normal user-end devices among all user-end devices associated with the vth edge node, is the transmission power of the mth normal user terminal device among all normal user terminal devices associated with the vth edge node at the tth iteration, is the channel gain from the mth normal user terminal device to the vth edge node at the tth iteration, is the local gradient standard value of the mth normal user terminal device among all normal user terminal devices associated with the vth edge node at the tth iteration, is the number of Byzantine attackers in all client devices associated with the vth edge node, At the tth iteration, The channel gain from a Byzantine attacker to the vth edge node is, is the number of Byzantine attackers associated with the vth edge node at the tth iteration. The local gradient of a Byzantine attacker, is the number of Byzantine attackers associated with the vth edge node at the tth iteration. The transmission power of a Byzantine attacker, is an additive Gaussian white noise that is independent of the gradient update; The cloud center server is used to broadcast the global model to each user terminal device via the edge node; it is also used to aggregate the gradient estimation values ​​of each edge node to update the global model, so as to obtain the optimal global model.

6. The cloud-edge service platform according to claim 5, characterized in that: The aggregation method adopted by the cloud center server is the geometric median aggregation method.

7. A distributed secure computing device based on a cloud-edge-end service platform, characterized in that: include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory to implement the method according to any one of claims 1 to 4.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 4 when executed by a processor.

Images