Secure communication and resource optimization method and system based on adaptive federated learning

Abstract

The invention belongs to the technical field of wireless communication and federated learning, and discloses a secure communication and resource optimization method based on adaptive federated learning, which realizes secure transmission with low complexity and low resource overhead by using the randomness of a wireless channel and the channel quality difference between a target channel and an eavesdropping channel through a physical layer security technology. Meanwhile, the quantization bit number is dynamically adjusted in combination with an adaptive quantization technology, and the parameter transmission quantity is remarkably reduced on the premise of ensuring the model precision, so that the gradient transmission delay is reduced. The invention further provides a power distribution algorithm, the secrecy rate of the minimum equipment is maximized by sending artificial noise to an eavesdropper, and the system performance is further optimized. A simulation experiment verifies that the security and the transmission efficiency of the federated learning system can be effectively improved, the uploading time delay is remarkably reduced while the model precision is ensured, and the method is particularly suitable for bandwidth-limited scenes such as satellite communication and the Internet of Things (IoT).

Description

Technical Field

[0001] The present invention belongs to the field of wireless mobile communication technology and relates to the research on non-orthogonal multiple access physical layer security in distributed Internet of Things. It uses a jammer to assist in increasing the uplink confidentiality rate and uses quantization technology to reduce the amount of uploaded parameters, thereby reducing the uplink transmission delay of federated learning. Background Art

[0002] Currently, satellite IoT can leverage satellites' wide-area coverage and long-range connectivity capabilities to provide efficient IoT services in remote areas, such as those difficult to reach with terrestrial communications. With the growing demand for data privacy and security, the implementation of new data privacy regulations has made access to distributed private data a major challenge for cross-node collaborative training. Traditional machine learning requires transmitting local raw data to a central server for centralized processing, but this process can expose users to the risk of information leakage. To address this issue, distributed machine learning, particularly federated learning, avoids centralized transmission through local training and aggregation, ensuring data privacy while reducing bandwidth and computing costs. This approach is suitable for scenarios with high privacy requirements, such as smart devices and the Internet of Things.

[0003] However, multiple factors must be considered when deploying a federated learning framework. First, while federated learning can protect a device's local data from direct leakage, during the process of uploading model parameters to the satellite and broadcasting updated parameters via satellite, an eavesdropper could potentially infer and restore local data through eavesdropped parameter information, posing certain data privacy risks. Second, the federated learning framework requires frequent local model updates, parameter uploads, and global aggregation, placing high demands on the satellite IoT system's communication, computing, and energy resources. However, SIoT resources are typically limited, making it difficult to support high-intensity local computing and data transmission.

[0004] To address data leakage in federated learning (FL), researchers have mostly approached data encryption from a cryptographic perspective. For example, they apply adaptive differential privacy models to federated learning frameworks to prevent inference attacks on endpoint data privacy, apply homomorphic encryption to federated learning to protect model parameters, and utilize multi-party computation to eliminate indirect leakage of local information by servers in federated learning. However, these algorithms require complex encryption algorithms or secure computations, resulting in high resource consumption. Therefore, it is crucial to develop a security technology with low complexity and resource consumption.

[0005] To address the problem of large data transmission volume, researchers have used quantization and sparsification technologies to reduce the amount of data transmission without significantly reducing the accuracy.

[0006] Through the above analysis, the problems and defects of the existing technology are as follows:

[0007] (1) Traditional upper-layer encryption technologies such as cryptography have high execution complexity and large resource overhead, and are not suitable for scenarios with limited communication resources such as satellite IoT.

[0008] (2) Most existing quantization technologies use fixed quantization algorithms and do not dynamically modify the number of quantization bits according to changes in the loss function, which is not flexible enough. Summary of the Invention

[0009] To address the challenges of existing technologies, we propose utilizing physical layer security to address the high complexity and resource overhead of traditional encryption techniques. From an information-theoretic perspective, physical layer security leverages the randomness of wireless channels and the difference in channel quality between the target and eavesdropped channels to achieve secure information transmission. Therefore, it offers significant advantages in terms of low complexity and resource overhead. Within the context of collaborative secure transmission using physical layer security, collaborative jamming techniques require only statistical channel state information from the eavesdropping node to degrade the received signal quality of the eavesdropping node by sending artificial noise to the eavesdropper. Without requiring complex encryption algorithms or security computations, basic security defenses can be implemented at low cost. Furthermore, the system's communication volume does not increase, making it suitable for bandwidth-constrained scenarios in satellite communications and the IoT. Furthermore, jammer deployment and operation are relatively simple, requiring only the identification of suitable jamming equipment. Collaborative transmission has a minimal impact on model performance. The jammer directly interferes with the communication channel and does not participate directly in the model update computation process. This helps maintain the accuracy of the federated learning model while ensuring data security.

[0010] Adaptive quantization technology is used to solve the problem of long delay caused by large amounts of data transmitted in federated learning models.

[0011] The present invention is implemented in the following way: through physical layer security technology, the randomness of the wireless channel and the difference in channel quality between the target channel and the eavesdropping channel are utilized to achieve low-complexity, low-resource-overhead secure transmission; adaptive quantization technology is used to dynamically adjust the number of quantization bits, reducing the parameter transmission amount while ensuring model accuracy, thereby reducing the gradient transmission delay; jammers are selected from devices not participating in federated learning, and artificial noise is sent to the eavesdropper to maximize the confidentiality rate of the minimum device and optimize system performance; power is allocated according to the channel gain of the devices participating in federated learning in the cluster and the selected jammer to maximize the confidentiality rate and minimize the upload delay.

[0012] Furthermore, the secure communication and resource optimization method based on adaptive federated learning further includes: during the federated learning process, the federated learning central server first broadcasts its gradients to the devices participating in the training. After local training, the devices upload the gradients to the central server, and the central server performs an aggregation operation. During the process of uploading gradients, eavesdroppers may restore the network parameters by eavesdropping on the gradient information. Therefore, an optimal edge device is selected from the set of devices that do not participate in the training. This device uses its own power to send artificial interference to the eavesdropper to prevent eavesdropping by the eavesdropper and increase the secrecy rate. The quantization technology is adopted in the uploading stage to reduce the amount of parameter transmission, thereby minimizing the gradient transmission delay.

[0013] Furthermore, the secure communication and resource optimization method based on adaptive federated learning includes the following steps:

[0014] Step 1, construct a network model and define the interaction model among multiple edge Internet of Things devices, base stations, and satellites;

[0015] Step 2, maximize the secrecy rate of the minimum device through the power allocation algorithm assisted by a jammer to optimize the system performance.

[0016] Step 3, adopt an adaptive quantization technology to dynamically adjust the quantization bit number to reduce the amount of parameter transmission;

[0017] Furthermore, the construction of the network model in Step 1 includes:

[0018] The edge devices participating in federated learning communicate with the base station. The global model parameters are obtained through distributed training. There are multiple edge Internet of Things devices N = {1,..., n,..., N} in the model, and the data set contained in each device itself is D = {D1,..., D N}, the devices participating in federated learning form a NOMA cluster M = {1,..., m,..., M}, M < N, M ∈ N, and the remaining set of devices that can act as jammers is R = {1,..., r,..., R}, R = N / M, the number of interactions with the base station is I, and the data volume in the i-th round of interaction is S i Furthermore, the model construction in Step 1 also includes:

[0019] One round of interaction requires four processes, which are divided into local training Upload parameters Aggregate parameters t agg And parameter broadcast t down Suppose that in the parameter broadcast stage, the base station can send interference to the eavesdropper while broadcasting the parameters, or in the broadcast stage, all the devices are idle at this time, so there are enough devices to send interference noise to the eavesdropper, resulting in a very low eavesdropping rate for the eavesdropper. Therefore, it is assumed that the secrecy rate in the broadcast stage is 2 bit / s / Hz. Therefore, the delay for one round The total delay of round I is ∑ i∈I T i .

[0020] Local training phase Local training energy consumption Therefore, the delay of the local training phase is The upload device set in the upload phase is M = {1,…,m,…M}, and the device upload phase delay is in The system upload delay is The local training energy consumption of device m is The aggregation phase delay is t agg , the confidentiality rate in the broadcast phase is The broadcast phase delay is

[0021] In the uplink phase, a device that does not participate in federated learning can be selected to act as a jammer to send artificial noise. The interference sent by the device can be decoded by the base station, that is, it has no impact on the base station but has an impact on the eavesdropper. Then the confidentiality rate in the system upload phase is can be rewritten as

[0022] The goal of the embodiment of the invention is to minimize the system communication delay by optimizing power allocation. The optimization problem can be expressed as:

[0023]

[0024] sth1 <h2<…<h M

[0025]

[0026] Since the optimization problem is the power allocation problem in the device upload phase, the local delay and energy consumption of the device can be obtained according to the simulation parameters. Therefore, the optimization problem can be transformed into

[0027]

[0028] sth1 <h2<…<h M

[0029]

[0030] Since the parameter amount uploaded in each round is a fixed value, the optimization problem is then transformed into

[0031]

[0032] sth1 <h2<…<h M

[0033]

[0034] Since the original problem is non-convex, the optimization problem is transformed into

[0035]

[0036] sth1 <h2<…<h M

[0037]

[0038] Now problem P4 itself is convex, and the constraint is also convex, so we need to and It is converted into a convex function through Taylor expansion.

[0039] for in

[0040] The problem then becomes

[0041]

[0042] Then converted into

[0043] Expand to

[0044]

[0045] We need to convert the left side of the equation into a convex function, first analyzing the first term

[0046] The partial derivative of p k The first-order partial derivative of

[0047]

[0048] P m The first-order partial derivative of

[0049]

[0050] Calculate the second-order partial derivatives based on the first-order partial derivative results, and get that all the second-order partial derivatives are positive, the function

[0051] f(p1,p2,…,p m ) is a convex function, and the same is true for other terms.

[0052] and It is a concave function and needs to be transformed into a convex function through Taylor expansion.

[0053]

[0054] against Can be converted into Convert to Convert to It is a fixed value. It has been proven above that m S i is a convex function, then the entire problem becomes a convex function, which can be solved using convex optimization tools.

[0055] Since the initial optimization goal of the present invention is to minimize latency, the present invention initially assumes that the amount of data transmitted in each round is a constant value in the upload latency, that is, no quantized gradient is performed, and each parameter is represented by 32 dimensions. However, many quantization schemes can now achieve higher accuracy using a low number of bits. Therefore, the present invention adopts an adaptive QSGD quantization scheme.

[0056] In the architecture of the present invention, there are M devices participating in federated learning, each of which has a data set D m , the labeled samples in the dataset are By parameter vector w∈R d To train the model, the optimization goal is to minimize the loss function

[0057]

[0058] in

[0059] The device uses the local gradient descent algorithm to iteratively train the model. At the beginning of each round i, each client obtains the global model w from the central server. i ,Then perform τ local updates during training, t=0,…,τ-1.

[0060]

[0061] After the local update is completed, the device uploads the gradient to the central server In order to reduce the amount of parameter transfer, quantization technology is used to convert the gradient into The central server updates based on the received parameters

[0062]

[0063] The present invention adopts a random uniform quantizer. The quantization level s∈N={1,2,…}, for the d-dimensional parameter vector w=(w1,w2,…,w d ],Q s (w i )=||w||2sign(w i )ζ i (w,s),

[0064]

[0065] l∈{0,1,2,…,s-1} is an integer. If w=0, then Q s (w)=0.

[0066] Given Q s (w i ) requires one bit to represent the sign, Indicates ζ i (w,s), the scalar ||w||2 is represented with full precision, so the number of bits transmitted from the client to the central server in each round is

[0067]

[0068] If the adaptive quantization scheme is used, the value of s in each round is

[0069]

[0070] Ultimately, the present invention's example obtains the optimal jammer within the cluster, and obtains the optimal confidentiality rate and adaptive quantization bits. Traditional confidential transmission does not utilize idle devices as jammers or randomly or fixedly selects jammers, so the confidentiality rate does not have the characteristic of maximization. The present invention's example proposes an optimal jammer selection algorithm, which increases the confidentiality rate of the devices participating in federated learning in the system, thereby reducing upload latency. The quantization bits of traditional quantization schemes are fixed and lack dynamic characteristics. The present invention's example uses an algorithm in which the quantization bits dynamically change according to the loss function, minimizing the number of uploaded parameters within the accuracy and loss function tolerance, thereby reducing upload latency.

[0071] Another object of the present invention is to provide a system for applying the secure communication and resource optimization method based on adaptive federated learning, the system comprising:

[0072] The physical layer security module is used to exploit the randomness of the wireless channel and the difference in channel quality between the target channel and the eavesdropped channel to achieve low-complexity and low-resource-overhead secure transmission;

[0073] Jammer auxiliary module, used to maximize the secrecy rate of the smallest device and optimize system performance by sending artificial noise to the eavesdropper;

[0074] The power allocation optimization module is used to allocate power based on the channel gains of the federated learning devices in the cluster and the selected jammers to maximize the confidentiality rate and minimize the upload delay.

[0075] Adaptive quantization module, used to dynamically adjust the number of quantization bits, reducing the amount of parameter transmission while ensuring model accuracy, thereby reducing gradient transmission delay;

[0076] Another object of the present invention is to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the steps of the federated learning secure transmission and resource optimization method based on physical layer security and adaptive quantization are performed.

[0077] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to perform the steps of the method for secure transmission and resource optimization of federated learning based on physical layer security and adaptive quantization.

[0078] Another object of the present invention is to provide an information data processing terminal, which is used to implement the steps of the federated learning secure transmission and resource optimization method based on physical layer security and adaptive quantization.

[0079] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0080] This invention addresses the high complexity and resource overhead of traditional encryption techniques in the existing art by proposing a solution that utilizes physical layer security technology to achieve secure information transmission. Based on information theory, physical layer security technology leverages the randomness of wireless channels and the difference in channel quality between the target channel and the eavesdropped channel to achieve secure transmission, offering significant advantages of low complexity and low resource overhead. Specifically, in collaborative secure transmission, collaborative jamming technology only requires eavesdropping on the statistical channel state information of the eavesdropping node and sending artificial noise to the eavesdropper to reduce the quality of its received signal. This eliminates the need for complex encryption algorithms or security calculations, enabling basic security protection at a low cost. This technical solution not only reduces the system's communication complexity but also avoids an increase in communication volume, making it particularly suitable for bandwidth-constrained scenarios such as satellite communications and the Internet of Things (IoT). Furthermore, the jammer is simple to deploy and operate, requiring only the selection of appropriate jamming equipment. It also has minimal impact on the performance of the federated learning model, as the jammer only ensures data security by interfering with the communication channel and does not directly participate in the model update computational process. This allows for efficient security protection while maintaining model accuracy.

[0081] To address the latency issues associated with large data transmission volumes in federated learning models, this paper proposes adaptive quantization technology. By dynamically adjusting the number of quantization bits, this technology significantly reduces the amount of parameter transmission while maintaining model accuracy, effectively reducing gradient transmission latency. Experimental results demonstrate that the adaptive quantization scheme outperforms the fixed quantization scheme in both convergence speed and final accuracy, particularly in scenarios with large data volumes.

[0082] Based on the main problems of the background technology and the limitations of the existing solutions, the present invention proposes a jammer-assisted power allocation algorithm, which aims to maximize the confidentiality rate of the minimum device through quantization algorithm and jammer assistance, thereby minimizing the upload delay.

[0083] The technical solution of the present invention significantly reduces system complexity and resource overhead by combining physical layer security technology with adaptive quantization technology, while also improving the security and efficiency of data transmission. This technical solution has broad application prospects in fields such as satellite communications and the Internet of Things (IoT), and can bring significant economic benefits to related industries. For example, in satellite communications, this solution can reduce communication latency and costs, and improve communication quality; in the IoT, this solution can enhance secure communication capabilities between devices and reduce deployment and maintenance costs. Therefore, the technical solution of the present invention has high commercial value and market potential.

[0084] The application of traditional encryption technologies in federated learning faces challenges such as high complexity, high resource consumption, and long latency. The introduction of physical layer security technology offers a new approach to addressing these challenges. This invention successfully addresses these challenges by combining physical layer security technology with adaptive quantization, achieving efficient, low-latency secure transmission and representing a significant technological breakthrough.

[0085] Targeting federated learning wireless transmission scenarios, this paper constructs a communication network model consisting of edge IoT devices, jammers, base stations, and eavesdroppers. This model assumes the presence of multiple edge devices, some of which participate in federated learning to form a NOMA cluster; the remaining devices are selected as jammers, injecting artificial noise signals into the eavesdropping channel during the federated learning upload phase. By defining a single federated learning interaction cycle as consisting of four phases: local training, quantized parameter upload, parameter aggregation, and parameter broadcast, and establishing precise expressions for latency and energy consumption for each phase, this model forms a joint optimization problem for minimizing overall federated learning communication latency and maximizing device confidentiality rates.

[0086] Traditional federated learning architectures typically fail to account for dynamic changes in wireless channel conditions and the impact of external eavesdropping threats during gradient upload and parameter interaction, leading to problems such as excessive transmission latency, insufficient confidentiality, and irrational resource allocation. Furthermore, existing fixed-bit quantization schemes incur significant communication overhead during parameter transmission and fail to dynamically adjust quantization accuracy based on the model's convergence characteristics, further increasing the network resource burden during federated learning training.

[0087] In response to the above technical problems, the present invention introduces physical layer security technology. Based on the inherent randomness of the wireless channel itself and the difference in channel quality between the target link and the eavesdropping link, it dynamically selects interference devices from the edge device set and effectively reduces the transmission capacity of the eavesdropping link through an artificial noise injection mechanism, thereby significantly improving the physical layer confidentiality performance of data transmission in the federated learning process, and providing a low-complexity security mechanism that does not require additional computing resources.

[0088] This paper also proposes an adaptive randomized uniform quantization (QSGD) scheme based on the convergence trend of the federated learning model. By analyzing the relationship between the model loss value of the current training round and the initial loss value, the number of quantization bits is dynamically adjusted, effectively reducing the amount of data transmitted for the gradient parameters during the federated learning process. This method can significantly reduce wireless communication resource overhead and effectively shorten the upload latency of model parameters while ensuring model convergence and accuracy, thereby further improving the transmission efficiency and scalability of the overall federated learning system. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0090] Figure 1 This is a flow chart of a secure communication and resource optimization method based on adaptive federated learning provided by an embodiment of the present invention;

[0091] Figure 2 Schematic diagram of a non-orthogonal multiple access system for edge IoT nodes provided by an embodiment of the present invention;

[0092] Figure 3 Schematic diagram showing the effect of applying the quantization technology provided by an embodiment of the present invention on the model accuracy when applied to the model gradient;

[0093] Figure 4 Schematic diagram showing the effect of applying the quantization technology provided by an embodiment of the present invention to the model gradient on the model loss function;

[0094] Figure 5 This is a schematic diagram of the adaptive bit number change of the adaptive quantization technology provided by an embodiment of the present invention;

[0095] Figure 6 Schematic diagram of minimum confidentiality rates obtained by adopting different schemes at different maximum receiving powers provided by an embodiment of the present invention;

[0096] Figure 7 This is a schematic diagram of the upload delay caused by different solutions as the amount of communication data increases, provided by an embodiment of the present invention.

[0097] Figure 8 This is a schematic diagram of the delay of different solutions provided by an embodiment of the present invention under the same model gradient but different quantization bits. DETAILED DESCRIPTION

[0098] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0099] In response to the problems existing in the prior art, the present invention provides an idea of using a jammer to ensure the security of uploaded parameters, and uses quantization technology to reduce model uploaded parameters. The present invention is described in detail below with reference to the accompanying drawings.

[0100] like Figure 1 As shown, an embodiment of the present invention provides a secure communication and resource optimization method based on adaptive federated learning, characterized by comprising the following steps:

[0101] S101: Using a physical layer security method to exploit the random characteristics of the wireless channel and the quality difference between the target channel and the eavesdropping channel, at least one edge device that does not participate in federated learning is selected as a jammer to inject an artificial noise signal into the eavesdropping channel.

[0102] S102, power is allocated based on the channel state information of the devices participating in the federated learning and the jammer to maximize the confidentiality rate of the minimum device;

[0103] S103: Adaptive random uniform quantization technology is used to dynamically adjust the number of quantization bits of the gradient parameters uploaded by the federated learning device to reduce the amount of uploaded parameter data.

[0104] As a preferred embodiment, the non-orthogonal multiple access physical layer security system for edge IoT nodes provided by the embodiment of the present invention specifically includes the following steps:

[0105] 1. Non-orthogonal multiple access physical layer security system for edge IoT nodes

[0106] In the construction of the network model, as Figure 2 shown, the embodiment of the present invention considers a model for edge Internet of Things, in which edge Internet of Things devices perform local training, and the base station obtains model parameters with high accuracy through multiple interaction rounds with them. The base station sends the parameters to the satellite, and the satellite sends the parameters to other remote ground devices. The specific parameter definitions are as follows: the total device set is N = {1,..., n,..., N}, the data set contained in each device itself is D = {D1,..., D N}, the devices participating in federated learning form a NOMA cluster, M = {1,..., m,..., M}, M < N, M ∈ N, and the remaining device set that can act as a jammer is R = {1,..., r,..., R}, R = N / M, the number of interaction rounds with the base station is I, and the data volume in the i-th round of interaction is S i , and one round of interaction requires four processes, which are divided into local training Upload parameters Aggregate parameters t agg and parameter broadcast t down . Assume that the base station can send interference to eavesdroppers while broadcasting parameters during the parameter broadcast phase, or during the broadcast phase, all devices are idle at this time, so there are enough devices to send interference noise to eavesdroppers, resulting in a very low eavesdropping rate for eavesdroppers. Therefore, assume that the secrecy rate during the broadcast phase is 2 bit / s / Hz. Therefore, the delay for one round

[0107]

[0108] Then the delay for the total number of rounds I is ∑ i∈I T i .

[0109] 1.1 Local training phase

[0110] where γ is the number of local training iterations, ε is the CPU cycles for training a single bit, v m is the local CPU rate. Therefore, the local training energy consumption τ is the effective capacitance switch. Therefore, the delay for the local training phase is

[0111] 1.2 Upload phase

[0112] The upload device set is M = {1,..., m,..., M}, and its corresponding channel gains are h1,..., h m ,..., h M , and satisfy h1 < h2 <... < h M , and the power assigned to each device is p1, p2,..., P M , and the total signal uploaded by the devices to the base station is

[0113]

[0114] The upload delay of device m is in is the confidential upload rate of device m,

[0115]

[0116] σ 2 and is Gaussian white noise, Latency of the upload phase of device m The system upload delay is The local training energy consumption of device m is

[0117] 1.3 Aggregation and Broadcasting

[0118] The aggregation phase delay is t agg , the confidentiality rate in the broadcast phase is The broadcast phase delay is

[0119] 1.4 Jammer Selection

[0120] In the uplink phase, a device that does not participate in federated learning can be selected to act as a jammer to send artificial noise. The interference sent by the device can be decoded by the base station, that is, it has no impact on the base station but has an impact on the eavesdropper. Then the confidentiality rate in the system upload phase is can be rewritten as

[0121]

[0122] The present invention assumes that the power of all idle devices is the same, that is, p1=p2=…=pR, then the optimal jammer is max{|h jame | 2}.

[0123] 2. Optimal power allocation

[0124] The goal of the embodiment of the invention is to minimize the system communication delay by optimizing power allocation. The optimization problem can be expressed as:

[0125]

[0126] sth1 <h2<…<h M (7a)

[0127]

[0128] Since the optimization problem is the power allocation problem in the device upload phase, the local delay and energy consumption of the device can be obtained according to the simulation parameters. Therefore, the optimization problem can be transformed into

[0129]

[0130] sth1 <h2<…<h M (8a)

[0131]

[0132] Since the parameter amount uploaded in each round is a fixed value, the optimization problem is then transformed into

[0133]

[0134] st(8a)(8b)(8c)

[0135] Since the original problem is non-convex, I transform the optimization problem into

[0136]

[0137] st(8a)(8b)(8c)

[0138]

[0139] Now problem P4 itself is convex, constraint (8b) is also convex, and (8c)(10a) needs to be converted into convex functions.

[0140] For (10a) in The problem then becomes

[0141]

[0142] Then it is converted into:

[0143]

[0144] Expand the function to:

[0145]

[0146] We need to convert the left side of the equation into a convex function, first analyzing the first term The partial derivative of p k The first-order partial derivative of

[0147]

[0148] P m The first-order partial derivative of

[0149]

[0150] Calculate the second-order partial derivative based on the first-order partial derivative result

[0151]

[0152] Therefore, all second-order partial derivatives are positive, and the function f(p1,p2,…,p m ) is a convex function, and the same is true for other terms. and It is a concave function and needs to be transformed into a convex function through Taylor expansion.

[0153]

[0154] Target (8c) Can be converted into Convert to Convert to It is a fixed value. It has been proven above that m S i is a convex function, then the entire problem becomes a convex function, which can be solved using convex optimization tools.

[0155] 3. Adaptive Quantization Scheme

[0156] Since the initial optimization goal of the present invention is to minimize latency, the present invention initially assumes that the amount of data transmitted in each round is a constant value in the upload latency, that is, no quantized gradient is performed, and each parameter is represented by 32 dimensions. However, many quantization schemes can now achieve higher accuracy using a low number of bits. Therefore, the present invention adopts an adaptive QSGD quantization scheme.

[0157] In the architecture of the present invention, there are M devices participating in federated learning, each of which has a data set D m , the labeled samples in the dataset are By parameter vector w∈R d To train the model, the optimization goal is to minimize the loss function

[0158]

[0159] in

[0160] The device uses the local gradient descent algorithm to iteratively train the model. At the beginning of each round i, each client obtains the global model w from the central server. i,Then perform τ local updates during training, t=0,…,τ-1.

[0161]

[0162] After the local update is completed, the device uploads the gradient to the central server In order to reduce the amount of parameter transfer, quantization technology is used to convert the gradient into The central server updates based on the received parameters

[0163]

[0164] The present invention adopts a random uniform quantizer. The quantization level s∈N={1,2,…}, for the d-dimensional parameter vector w=[w1,w2,…,w d ],Q s (w i )=||w||2sign(w i )ζ i (w,s),

[0165]

[0166] l∈{0,1,2,…,s-1} is an integer. If w=0, then Q s (w)=0.

[0167] Given Q s (w i ) requires one bit to represent the sign, Indicates ζ i (w,s), the scalar ||w||2 is represented with full precision, so the number of bits transmitted from the client to the central server in each round is

[0168]

[0169] If the adaptive quantization scheme is used, the value of s in each round is

[0170]

[0171] Ultimately, the present invention's example obtains the optimal jammer within the cluster, and obtains the optimal confidentiality rate and adaptive quantization bits. Traditional confidential transmission does not utilize idle devices as jammers or randomly or fixedly selects jammers, so the confidentiality rate does not have the characteristic of maximization. The present invention's example proposes an optimal jammer selection algorithm, which increases the confidentiality rate of the devices participating in federated learning in the system, thereby reducing upload latency. The quantization bits of traditional quantization schemes are fixed and lack dynamic characteristics. The present invention's example uses an algorithm in which the quantization bits dynamically change according to the loss function, minimizing the number of uploaded parameters within the accuracy and loss function tolerance, thereby reducing upload latency.

[0172] This application example illustrates the application of the non-orthogonal multiple access physical layer security research algorithm proposed in the present invention to a distributed Internet of Things. By optimizing power allocation, jammer selection, and quantization schemes, the uplink transmission delay problem of the distributed Internet of Things can be solved, thereby reducing transmission delay and improving system efficiency.

[0173] The performance of the proposed solution was verified through computer simulation. The experiment used the MNIST dataset, and the local datasets of each device participating in federated learning met the independent and identically distributed condition. Both system noise and path loss were normalized.

[0174] Figure 3 The paper shows how the model accuracy changes with the number of training rounds under different quantization bit numbers. Experimental results show that all four quantization schemes reach convergence when the data volume is 0.25×10^7 bits. Among them, the larger the number of quantization bits, the slower the convergence speed; it is worth noting that when the number of quantization bits is 4 bits, its final accuracy is significantly lower than that of the other three quantization schemes. In contrast, the adaptive quantization scheme proposed in this paper can not only guarantee the model accuracy, but also achieve faster convergence speed, thereby effectively improving training efficiency.

[0175] Figure 4 The loss function changes with the number of training rounds under different quantization bits, and its convergence characteristics are similar to Figure 2 The accuracy change trends shown are consistent, which further verifies the effectiveness of the proposed adaptive quantization scheme.

[0176] Figure 5 This paper demonstrates the dynamic adjustment of the number of quantization bits in the adaptive quantization algorithm over training rounds. Initially, the system uses a smaller number of quantization bits. As the loss function decreases, the number of quantization bits increases, eventually stabilizing at 7 bits, according to the adaptive calculation formula.

[0177] Figure 6The results demonstrate the performance and confidentiality rates of various secure communication schemes. Experimental data demonstrates that the optimal jammer selection scheme proposed in this paper significantly outperforms other schemes. Notably, the algorithm that utilizes idle devices as jammers achieves a higher confidentiality rate than the scheme that does not. This result fully demonstrates the effectiveness of using physical layer security technologies to improve system confidentiality rates.

[0178] Figure 7 The paper demonstrates how system upload latency changes with increasing data volume using different jammer selection algorithms. Experimental results show that, with small data volumes, the performance differences between the algorithms are minimal; however, as data volumes increase, these differences become more pronounced. The advantages of the optimal jammer selection algorithm are particularly pronounced in large data volumes.

[0179] Figure 8 The data volume changes of different secure communication schemes under different quantization bit numbers were compared. Experimental data showed that the system delay was minimized when the adaptive quantization scheme was adopted, further verifying the effectiveness of the scheme of the present invention.

[0180] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0181] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A secure communication and resource optimization method based on adaptive federated learning, characterized in that: The steps include: Step 1: Using physical layer security methods to exploit the random characteristics of the wireless channel and the quality difference between the target channel and the eavesdropping channel, at least one edge device that does not participate in federated learning is selected as a jammer to inject an artificial noise signal into the eavesdropping channel. Step 2: Power is allocated based on the channel state information of the devices participating in federated learning and the jammer to maximize the confidentiality rate of the minimum device; Step 3: Adaptive random uniform quantization technology is used to dynamically adjust the number of quantization bits of the gradient parameters uploaded by the federated learning device to reduce the amount of uploaded parameter data.

2. The secure communication and resource optimization method based on adaptive federated learning according to claim 1, characterized in that: The federated learning process includes: a central server broadcasts global model gradient parameters to participating devices; participating devices perform local gradient updates of model parameters based on local data sets; an adaptive random uniform quantizer is used to quantize local gradients; the quantized gradients are uploaded to the central server; and the central server aggregates and updates the received quantized gradient parameters.

3. The secure communication and resource optimization method based on adaptive federated learning according to claim 1, characterized in that: The network model includes multiple edge IoT devices. The devices participating in federated learning form a non-orthogonal multiple access (NOMA) cluster. The devices not participating in federated learning are used as jammers. The single iterative interaction process of federated learning is defined as including local model training, uploading quantized gradient parameters, server-side aggregation and model parameter broadcasting.

4. The secure communication and resource optimization method based on adaptive federated learning according to claim 3 is characterized in that: The delay of each round of interaction in the federated learning includes the device local training delay, parameter upload delay, server-side aggregation delay and server-side parameter broadcast delay; the device local training delay is determined by the number of device local training cycles, data set size and computing power; the device parameter upload delay is determined by the amount of uploaded data and the security rate, where the security rate in the upload stage is expressed as the link capacity between the device and the base station minus the average capacity of the link between the device and the eavesdropper.

5. The secure communication and resource optimization method based on adaptive federated learning according to claim 3, characterized in that: The power allocation optimization steps are as follows: among the devices participating in federated learning, the devices are sorted from low to high based on the channel gain between the devices and the base station, and a confidentiality rate optimization model is constructed with the transmission power of each device as a variable; with the total power of the devices and the energy consumption of the devices during the upload phase as constraints, the non-convex constraints are Taylor expanded to approximate convex functions, and then the optimal power allocation of each device is obtained using a convex optimization tool.

6. The secure communication and resource optimization method based on adaptive federated learning according to claim 1, characterized in that: The adaptive random uniform quantization technology includes: in each round of federated learning, each device dynamically adjusts the random uniform quantization level according to the current global model parameters and the local loss function value. The quantization level is proportional to the initial level. The quantizer performs probabilistic quantization on the gradient parameters and uploads the quantized gradients to the central server. The central server completes the model parameter aggregation update based on the received quantized gradients.

7. A system using the secure communication and resource optimization method based on adaptive federated learning according to any one of claims 1 to 6, characterized in that: The secure communication and resource optimization system based on adaptive federated learning includes: The physical layer security module is used to exploit the randomness of the wireless channel and the difference in channel quality between the target channel and the eavesdropped channel to achieve low-complexity and low-resource-overhead secure transmission; Jammer auxiliary module, used to maximize the secrecy rate of the smallest device and optimize system performance by sending artificial noise to the eavesdropper; The power allocation optimization module is used to allocate power based on the channel gains of the federated learning devices in the cluster and the selected jammers to maximize the confidentiality rate and minimize the upload delay. The adaptive quantization module is used to dynamically adjust the number of quantization bits, reducing the amount of parameter transmission while ensuring model accuracy, thereby reducing gradient transmission delay.

8. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the secure communication and resource optimization method based on adaptive federated learning according to any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the secure communication and resource optimization method based on adaptive federated learning according to any one of claims 1 to 6.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the secure communication and resource optimization system based on adaptive federated learning as described in claim 7.

Images