Semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing

Abstract

The application discloses a semantic communication privacy enhanced federated multi-task learning method of mobile edge computing, wherein a method comprises the following steps: an edge server initializes a global model comprising a shared encoder and multiple personalized decoders and distributes the global model; each mobile edge client performs multi-task training locally, and then performs integrated error compensation, sparsification, quantization and semantic compression processing on model updates, and uses a local private codebook to discretize and map features to enhance privacy, and then uploads compressed data; after receiving the data, the server end infers a non-linear relationship graph between global tasks through adaptive kernel learning, and performs structure-aware differentiated weighted aggregation on the updates of the shared encoder and each personalized decoder according to the non-linear relationship graph, and updates the global model and distributes the global model to the client for the next round of iteration. The application reduces communication overhead while realizing endogenous privacy protection, and effectively improves the efficiency, safety and performance of federated multi-task learning in a resource-limited mobile edge environment.

Description

Technical Field

[0001] This invention relates to the fields of federated learning, multi-task learning, and privacy protection, and particularly to a semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing. Background Technology

[0002] With the deep integration of mobile edge computing and artificial intelligence technologies, the application of edge nodes such as smart terminals, IoT sensors, embedded devices, mobile edge gateways, vehicle terminals, and edge base stations is becoming increasingly widespread. These devices often need to concurrently execute multiple heterogeneous tasks, such as semantic segmentation, depth estimation, scene understanding, intelligent perception, edge inference, and edge-edge collaborative decision-making, in resource-constrained, low-bandwidth, and high-latency environments of mobile edge computing. Moreover, the data is stored locally in a dispersed manner, posing strict privacy protection requirements. However, when training personalized models across devices, issues such as data silos, heterogeneous environment adaptation, and privacy leakage risks are intertwined, making traditional centralized learning insufficient to meet actual deployment needs.

[0003] Federated Multi-Task Learning (FMTL) has emerged as a new distributed learning paradigm. This paradigm allows multiple mobile edge clients to collaboratively train personalized models adapted to local tasks without sharing raw data. It breaks down data silos and ensures privacy and security, providing an effective technical path for multi-task collaboration in mobile edge intelligence scenarios. However, most existing FMTL methods are implicitly designed under the assumption of relatively sufficient computing and communication resources, which cannot adapt to the heterogeneous resources and dynamic network environment of mobile edge computing. This assumption is often violated in real-world mobile edge intelligence scenarios. In actual deployment, FMTL still faces fundamental constraints caused by the dual heterogeneity of task heterogeneity and system environment heterogeneity. Especially in scenarios where mobile edge computing devices have limited power, network topology changes dynamically, and edge node computing power varies significantly, the training latency, communication lag, and device overload problems of traditional methods are more prominent, seriously affecting their training performance and system efficiency.

[0004] To address the task heterogeneity problem in federated multi-task learning, numerous scholars have conducted targeted research. Smith et al. (V. Smith, C.-K. Chiang, M. Sanjabi, and ASTalwalkar, “Federated multi-task learning,” Advances in Neural Information Processing Systems, vol. 30, 2017.) proposed the classic federated multi-task learning framework MOCHA, employing a hard parameter sharing mechanism to decompose the model into cross-task shared representations and task-specific components, laying the foundation for subsequent FMTL research. However, this framework did not fully consider the complex nonlinear dependencies between tasks. Wu et al. (C. Wu, F. Wu, L. Lyu, Y. Huang, and X. Xie, “Communication-efficient federated learning via knowledge distillation,” Nature) Communications, vol.13, no.1, p.2032, 2022. proposed a soft parameter sharing method based on knowledge distillation, which achieves personalized federated learning through indirect knowledge transfer. It improves the synergistic effect while preserving model heterogeneity, but it relies on the assumption of linear knowledge transfer and is difficult to adapt to nonlinear task associations in high-dimensional scenarios. Fallah et al. (A. Fallah, A. Mokhtari, and A. Ozdaglar, “Personalized federated learning: A meta-learning approach,” arXiv preprint arXiv:2002.07948, 2020.) introduced the idea of ​​meta-learning to design a personalized federated learning scheme. It adapts to different mobile edge client tasks by learning general initialization parameters, but it lacks explicit modeling of task structure and is easily affected by differences in data distribution. While these methods have alleviated the challenges posed by task heterogeneity to some extent, they still suffer from low knowledge transfer efficiency and significant negative transfer in edge computing scenarios. They generally lack in-depth modeling of the nonlinear structure of the global task, resulting in a lack of targeted knowledge transfer. Redundant information from irrelevant tasks can easily trigger negative transfer, making it difficult to balance model generality and task specificity.

[0005] Research on communication efficiency optimization and privacy protection continues to advance. McMahan et al. (B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. Ay Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial Intelligence and Statistics, PMLR, 2017, pp. 1273–1282.) proposed the classic FedAvg algorithm, which achieves efficient federated learning through weighted average updates from mobile edge clients. However, its naive weighting strategy ignores the semantic importance of parameters, easily leading to the loss of key information, and lacks a dedicated privacy protection design. Li et al. (T. Li, AK Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V. Smith, “Federated optimization in heterogeneous networks,” Proceedings of Machine Learning and Systems, vol.2, pp.429–450, 2020.) proposes a federated optimization scheme for heterogeneous network scenarios, which improves system robustness by adapting to differences in device resources. However, it does not deeply mine the semantic features in the communication process, making it difficult to balance the communication overhead and model performance under the low bandwidth and high packet loss rate of mobile edge computing.

[0006] Current research has not yet formed a general framework that systematically integrates model updates and semantic representation of data features. This framework fails to meet the practical deployment requirements of edge-end collaboration, ubiquitous sensing, and real-time inference in mobile edge computing, and makes it difficult to achieve a harmonious balance between communication efficiency, privacy, and model performance in heterogeneous edge environments. Therefore, how to overcome the limitations of traditional linear modeling and construct an FMTL framework that integrates nonlinear task relationship modeling, structure-aware aggregation, semantic-level communication compression, and privacy enhancement to achieve a harmonious balance between efficient collaborative training, low communication overhead, and privacy under dual heterogeneity has become a key technical problem urgently needing to be solved in the field of mobile edge intelligence. Summary of the Invention

[0007] The purpose of this invention is to provide a semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing. By deeply integrating kernelized structure reasoning with privacy-enhanced semantic communication, it adapts to the resource-constrained, edge-heterogeneous, and dynamic network characteristics of mobile edge computing, and achieves efficient collaborative training and privacy protection under dual heterogeneity.

[0008] The technical solution adopted in this invention is:

[0009] A semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing is applied in a federated learning system with one edge server and multiple mobile edge clients. The method includes the following steps:

[0010] Model distribution: The edge server builds and initializes a global model, which includes a shared encoder network for extracting cross-task shared features and multiple personalized decoder networks corresponding to the local tasks of each mobile edge client, and distributes the global model to each mobile edge client.

[0011] Local training: Each mobile edge client trains the received global model based on the private heterogeneous multi-task dataset stored locally to generate local model updates. Local model updates include parameter updates for the shared encoder network and parameter updates for the personalized decoder network corresponding to the task.

[0012] Parameter compression: Each mobile edge client compresses the parameter semantic representation of the local model update, obtains the compressed parameter semantic representation, and uploads it to the edge server through the mobile edge network;

[0013] Relational reasoning: The edge server reconstructs the local model update based on the semantic representation of parameters uploaded by each mobile edge client, and infers the nonlinear global task relationship graph between tasks on all edge clients through the adaptive kernel learning method, thus obtaining the self-representation matrix;

[0014] Aggregation and synchronization: The edge server performs structure-aware two-layer aggregation on the shared encoder and personalized decoder based on the self-representation matrix to obtain the updated global model and distribute it to each mobile edge client; and repeats the local training, parameter compression, relation inference and aggregation and synchronization steps until the global model converges and meets the end-edge collaborative training performance requirements of mobile edge computing.

[0015] Furthermore, the implementation of parameter semantic representation compression includes:

[0016] A local private codebook is generated and maintained based on the semantic feature distribution of local multi-task data.

[0017] Error compensation and sparsification are performed on the local model update parameters to filter out important parameters and location information;

[0018] The selected important parameter values ​​are quantized and dequantized to reconstruct the residual vector, which is used for error compensation in the next round.

[0019] The location index sequence of location information is indexed, encoded, and then compressed to obtain a compressed index bitstream;

[0020] The quantized residual vector is used to map the local task features of the current round to the local private codebook to obtain the quantization index and update the quantized residual vector.

[0021] Furthermore, the implementation of generating the local private codebook includes: performing covariance analysis on the local multi-task semantic features of the edge client to obtain the effective intrinsic dimension, and adaptively determining the codebook size based on the effective intrinsic dimension and the number of tasks. The local private codebook is stored locally on the client and is not uploaded.

[0022] Furthermore, the implementation of location information compression encoding includes: sorting the parameter location index sequence, differential encoding, and applying an entropy encoding algorithm to generate a compressed bit stream.

[0023] Furthermore, the implementation of the adaptive kernel learning method includes: the edge server reconstructs the task representation matrix based on the semantic representation of parameters uploaded by each mobile edge client, and the task representation matrix obtains an initial representation matrix through preliminary linear self-representation learning; a symmetric affinity matrix is ​​constructed based on the initial representation matrix and normalized to obtain a relation matrix; an adaptive, task-driven kernel matrix is ​​constructed based on the relation matrix G; and an alternating optimization method is used to solve the optimization problem of the adaptive kernel matrix. After convergence, the self-representation matrix is ​​learned as a global task relation graph, and the self-representation matrix provides a structured representation for the global task manifold.

[0024] Furthermore, the implementation of two-level aggregation includes:

[0025] The first aggregation weight is calculated for each edge client based on the self-representation matrix, and the shared encoder network parameter updates from each edge client are weighted and aggregated with the first aggregation weight to update the global shared encoder.

[0026] For each task, a second aggregation weight is calculated for each edge client associated with the corresponding task based on the self-representation matrix. The second aggregation weight is then used to weight and aggregate the personalized decoder network parameter updates of the relevant tasks of the associated edge clients to update the global personalized decoder of each task.

[0027] Furthermore, calculating the first aggregation weight specifically includes: summing the absolute values ​​of the affinity between all tasks on each edge client and tasks on other edge clients in the self-representation matrix, and globally normalizing the cumulative values ​​of each edge client to obtain the weight of each edge client in the shared encoder aggregation.

[0028] Furthermore, calculating the second aggregation weight specifically includes: selecting edge clients from the self-representation matrix whose affinity on the specified task exceeds a preset threshold to form a set; calculating and normalizing the relative centrality score of each edge client in the set on the corresponding task; and obtaining the weight of each edge client in the personalized decoder aggregation of the corresponding task.

[0029] This invention, employing the above technical solutions, offers the following advantages compared to existing technologies: 1) The global nonlinear task relationship graph constructed through the adaptive structural reasoning module overcomes the limitations of traditional linear assumptions in complex real-world scenarios. This enables the system to more reliably identify the essential relationships between tasks in edge environments where network topology and device states dynamically change, providing accurate basis for subsequent resource allocation and model collaboration, and improving the system's decision robustness and reliability in real, unstable environments. 2) A structure-aware, two-layer aggregation strategy is adopted to implement differentiated fusion of the shared encoder and personalized decoder. This not only optimizes the utilization efficiency of limited computing and storage resources on the edge side and avoids redundant computation, but also ensures that when serving numerous heterogeneous devices, the global model possesses both the universality of common knowledge and the personalized adaptability to specific terminals or scenarios, thereby meeting the diverse Quality of Service (QoS) requirements of MEC. 3) An integrated dual semantic communication mechanism combines model and data compression with inherent privacy protection. This mechanism, through sparsification, quantization, and dedicated local codebook technologies, significantly reduces communication load and transmission latency in low-bandwidth, high-latency mobile edge networks. Meanwhile, its feature discretization and local codebook design effectively resist common edge-side data theft and gradient inversion attacks without relying on additional encryption overhead, providing feasible security technology guarantees for deploying federated learning in untrusted or open network environments. 4) Through adaptive codebook size and dynamic quantization strategies, this invention flexibly adapts to different types of edge devices, from ultra-low-power IoT sensors to high-performance edge gateways and in-vehicle mobile terminals. This invention can be efficiently deployed in real-world mobile edge computing scenarios with strictly limited resources and heterogeneous hardware, and maintains linear system scalability as the scale of edge nodes increases. Attached Figure Description

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;

[0031] Figure 1 This is a flowchart illustrating the semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing according to the present invention.

[0032] Figure 2 This is a schematic diagram illustrating the principle framework of the semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing of the present invention.

[0033] Figure 3 This is a schematic diagram illustrating the specific process of the semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing according to the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0035] This invention addresses the context of edge-to-edge collaborative multi-task training in heterogeneous mobile edge computing scenarios. It combines adaptive kernel learning and privacy-enhancing semantic communication techniques to construct the FedAtlas framework. This framework can simultaneously handle heterogeneous tasks across mobile edge clients and the low-bandwidth, high-latency, and resource-constrained mobile edge communication environment, achieving efficient collaborative training of multi-task models. Through a dual mechanism of kernelized structure inference and structure-aware aggregation, it accurately captures nonlinear dependencies between tasks, promoting positive knowledge transfer and suppressing negative interference. By combining a mobile edge client-specific codebook with multi-level semantic compression technology, it significantly reduces communication overhead while providing inherent privacy protection without additional noise injection, thereby improving the performance, efficiency, and security of federated multi-task learning in heterogeneous edge scenarios.

[0036] like Figures 1 to 3 As shown in any of the accompanying drawings, embodiments of the present invention disclose a semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing, comprising the following steps:

[0037] S1. The edge server builds and initializes a global model, which includes a shared encoder network for extracting cross-task shared features and multiple personalized decoder networks corresponding to the local tasks of each mobile edge client. The global model is then distributed to each mobile edge client.

[0038] S2. Local training: Each mobile edge client trains the received global model based on the private heterogeneous multi-task dataset stored locally to generate local model updates. Local model updates include parameter updates for the shared encoder network and parameter updates for the personalized decoder network corresponding to the task.

[0039] S3. Each mobile edge client compresses the parameter semantic representation of the local model update, obtains the compressed parameter semantic representation, and uploads it to the edge server through the mobile edge network.

[0040] S4. The edge server reconstructs the local model update based on the semantic representation of parameters uploaded by each mobile edge client, and infers the global task relationship graph through adaptive kernel learning to obtain the self-representation matrix.

[0041] S5. Based on the self-representation matrix, the edge server performs structure-aware two-layer aggregation on the shared encoder and personalized decoder to obtain an updated global model and distribute it to each mobile edge client.

[0042] S6. Repeat steps S2 to S5 until the model converges and meets the edge-to-edge collaborative training performance requirements of mobile edge computing.

[0043] In this embodiment, the deep fusion of adaptive kernel learning and semantic communication effectively solves the dual heterogeneity challenge of federated multi-task learning in heterogeneous mobile edge scenarios. The kernelized structure inference mechanism accurately characterizes the complex nonlinear dependencies between tasks, providing precise guidance for differentiated knowledge sharing. The structure-aware two-layer aggregation strategy optimizes the shared encoder and personalized decoder respectively, balancing model generality and task specificity. The combination of a dedicated codebook for the mobile edge client and multi-level semantic compression technology significantly reduces communication overhead while effectively resisting attacks such as gradient inversion through feature discretization and local codebook isolation. The entire framework achieves performance improvements in non-independent, identically distributed data scenarios without requiring additional privacy protection modules or high-performance hardware support, perfectly adapting to the collaborative training needs of heterogeneous mobile edge devices such as mobile terminals and IoT sensors, providing an efficient and secure solution for multi-task collaboration in edge intelligence scenarios.

[0044] Please refer to Figures 1 to 3 Based on the above embodiments, in step S1, the edge server initializes a global model, which consists of a shared encoder and multiple personalized decoders. The shared encoder is used to extract feature representations shared across tasks, while the personalized decoders are used to process the output of specific tasks. Each mobile edge client i possesses... Each local task uses a local dataset to perform multiple rounds of local training on the downloaded global model to obtain the current communication round of the mobile edge network. Local model update The local model update includes shared encoder updates. and personalized decoder updates , where t represents the t-th task.

[0045] Furthermore, in step S2, each mobile edge client trains the downloaded global model locally based on its local multi-task data. Each mobile edge client i uses its local dataset. The model is optimized through multiple rounds of gradient descent to obtain local model update parameters adapted to mobile edge computing.

[0046] In this embodiment, by distinguishing between cross-task shared features and task-specific outputs, the foundation for subsequent differentiated aggregation is laid. It emphasizes the multi-task characteristics and non-independent identically distributed attributes of local data, aligning with the actual data distribution characteristics of mobile edge intelligence scenarios; and introduces mobile edge network communication rounds. Labeling model updates ensures the orderliness of iterative training, provides temporal support for subsequent error compensation, semantic compression and other mechanisms, and effectively connects global collaboration with local personalized training needs.

[0047] Furthermore, the parameter semantic representation compression in step S3 above specifically includes the following steps:

[0048] S30. Mobile edge client-specific codebook construction: Each mobile edge client i maintains a private codebook as a unified abstract space for all local tasks, which is never transmitted to the edge server.

[0049] S31. Semantic Feature Extraction and Covariance Calculation: Let Let T represent the set of semantic features extracted from T local tasks of mobile edge client i, and calculate the empirical covariance matrix:

[0050] (1);

[0051] in, The mean of features is independent of the global task. Let be the semantic features of the t-th local task of mobile edge client i;

[0052] S32. Effective Intrinsic Dimension Estimation: Eigenvalue Decomposition of the Covariance Matrix Estimating the effective intrinsic dimension of the multi-task semantic space through effective rank estimation. :

[0053] (2);

[0054] in, To satisfy the minimum integer requirement of cumulative variance contribution rate ≥ 95%, For feature dimension, where is the eigenvalue of the covariance matrix.

[0055] S33. Adaptive Codebook Size Determination: Adaptively Determine Codebook Bit Width and codebook size :

[0056] (3);

[0057] (4);

[0058] For ultra-low power edge sensors, a smaller b is chosen to ensure device availability; for high-performance edge gateways, a larger codebook can be chosen to achieve higher fidelity.

[0059] S34. Codebook Initialization and Optimization: The k-means++ algorithm is used to initialize and optimize the joint multi-task feature set. Select M codewords as the initial codebook. The first codeword is sampled uniformly, and subsequent codewords are sampled uniformly for each currently selected codeword. Select based on the minimum squared distance from existing codewords:

[0060] (5);

[0061] (6);

[0062] in, For the first Distance from each task feature to the nearest selected codeword, The selected codeword sequence number and k <j, For the first Distance values ​​of each task feature, The codebook One initial codeword has been selected. This represents the total number of local tasks on the client side, in relative terms. Based on probability The currently selected task is obtained by sampling from the task feature set. The codebook is refined iteratively using standard k-means optimization with initial codewords: In the E-step, each task feature... Indexing is obtained by assigning to the most recent codeword In M steps, each codeword is updated to the mean of the task feature assigned to it. This iteration continues until the total displacement of all cluster centers satisfies:

[0063] (7);

[0064] in, Indicates the first The code character is in the first The value after round of iterations, Indicates the first The code character is in the first The value after round iteration; For codebook size, This is the convergence threshold.

[0065] S35. Error Compensation: Mobile edge client i performs error compensation on local model updates. The calculation formula is as follows:

[0066] (8);

[0067] in, For the current (r) round of local model update, The residual vector retained from the previous round, in the initial round. =0.

[0068] S36, Top-S sparsification: For the compensated vector... By applying Top-S sparsity techniques, the S elements with the largest absolute values ​​are extracted to obtain the position index sequence. and the corresponding parameter value sequence :

[0069] (9);

[0070] (10);

[0071] in, Let S represent the indices of the S elements with the largest absolute values, where S can be any value satisfying the condition... d represents the dimension of the model parameters.

[0072] S37. Quantization operation: Quantizing the parameter value sequence Quantize each element Mapping to a finite bit representation yields a quantized sequence. :

[0073] (11);

[0074] in, To quantize bit width, Represents a sparse value sequence Index of an element in the middle, This represents the quantization interval, where L is the lower limit of the quantization interval and R is the upper limit of the quantization interval.

[0075] S38. Dequantization and Reconstruction: To accurately calculate the quantization error for compensation in the next round, a dequantization operation is performed to obtain the reconstructed value.

[0076] (12);

[0077] Based on reconstructed value sequence and index Construct sparse update vector Fill the unselected positions with zeros.

[0078] S39. Residual Update: Calculate the residual vector for the next round of error compensation.

[0079] (13);

[0080] S310, Index Encoding: For position index sequences Differential coding and variable-length coding are used to further reduce communication overhead. First, the index set... Sort in ascending order, first index Use length is Fixed-length bit string encoding, where, This represents the total dimension of the model parameters; for each subsequent index, only the difference between it and the previous index is transmitted:

[0081] (14);

[0082] in, For the sorted number One index value, The difference between adjacent indices. The number of parameters retained after sparsification.

[0083] Huffman coding is used for the difference sequence Compression is performed by mapping high-frequency differences into shorter binary codes according to a pre-agreed Huffman code, generating the final compressed index bitstream. .

[0084] S311. Data semantic representation compression: An error feedback mechanism is used to quantize and compress task features.

[0085] (15);

[0086] (16);

[0087] (17);

[0088] in, For the first Wheel client Task The original semantic features, For codebook size, For the first The first round Each code character For index The corresponding codeword vector, These are the features after error compensation. For quantization index, This is the quantization residual for the current round. The mobile edge client only transmits discrete indices. To the edge server, codebook Stored locally and never transmitted, providing inherent structured privacy protection.

[0089] In this embodiment, a complete semantic communication link is constructed: the mobile edge client's dedicated codebook achieves accurate adaptation to the multi-task semantic space through adaptive dimensionality estimation and iterative optimization; the parameter semantic representation significantly reduces communication overhead while retaining key update information through multi-level compression of "error compensation-sparseness-quantization-index encoding"; and the data semantic representation avoids the leakage of original features and achieves inherent privacy protection through local codebook mapping and the design of transmitting only discrete indices. The entire process balances communication efficiency and data security, perfectly adapts to the bandwidth and privacy constraints of mobile edge devices, and solves the problems of traditional compression technologies ignoring semantic importance and the difficulty in balancing privacy protection and performance.

[0090] Furthermore, the adaptive kernel learning inference of the global task relationship graph in step S4 above includes the following specific steps:

[0091] S40. Task Representation Matrix Construction: The edge server collects the semantic representations of parameters uploaded by each mobile edge client and reconstructs the task representation matrix. ;in, Let N represent the feature representation of the t-th task of mobile edge client i, where N is the number of mobile edge clients and T is the number of tasks for each mobile edge client.

[0092] S41. Preliminary Linear Self-Representation Learning: Obtaining the initial representation matrix through preliminary linear self-representation learning. The formula is:

[0093] (18);

[0094] in, This is a tradeoff parameter for the regularization term, used to control the strength of regularization. For regularization terms, This is an operator for extracting the elements of the main diagonal of a matrix.

[0095] S42. Construction of Symmetric Affinity Matrix: Based on Initial Matrix Constructing a symmetric affinity matrix :

[0096] (19);

[0097] S43. Relationship matrix normalization: [This part is incomplete and requires further context.] Normalization is performed to obtain the relation matrix G, which enhances the cluster structure and suppresses the influence of high-order nodes:

[0098] (20);

[0099] in, Represented by vector It is a diagonal matrix composed of diagonal elements. It is a vector consisting entirely of 1s.

[0100] S44. Adaptive Kernel Matrix Construction: Based on the relation matrix G, construct an adaptive, task-driven kernel matrix K, defined as:

[0101] (twenty one);

[0102] in, and They represent the client respectively. Task and client Task , Relationship matrix The value of the element at the corresponding position in the middle. , For summation index variables, , It is a small constant to ensure the positive definiteness of K.

[0103] S45. Self-representation matrix optimization: Solving the optimization problem based on the adaptive kernel matrix K to obtain the self-representation matrix Z:

[0104] (twenty two);

[0105] in, For matrix trace operations, Q is an auxiliary variable used to enforce the symmetry and nonnegativity of Z, and P is used for non-convex diagonal regularization. The positive trade-off parameter.

[0106] S46. Alternating Optimization Solution: Solve the above non-joint convex optimization problem using the alternating optimization method, repeating the following three main steps until convergence:

[0107] First, update Z: fix Q and P, and solve the subproblems of Z:

[0108] (twenty three);

[0109] This problem has a closed-ended solution:

[0110] (twenty four);

[0111] in, To alternately optimize the number of iterations, It is an identity matrix.

[0112] Then, update P: fix Z and Q, and solve the subproblems of P:

[0113] (25);

[0114] This solution is obtained by... The terms are obtained by performing eigenvalue decomposition.

[0115] Finally, update Q: fix Z and P, and solve the subproblems of Q:

[0116] (26);

[0117] This problem also has a closed-form solution.

[0118] S47. Task Relationship Graph Generation: After convergence, the learned self-representation matrix Z provides a structured representation of the global task manifold. Each entry By capturing the nonlinear pairwise affinity between tasks (i,t) and (j,s), the induced block diagonal pattern reveals potential task groupings, providing precise task relationship guidance for subsequent structure-aware aggregation.

[0119] In this embodiment, adaptive kernel learning breaks through the limitations of traditional linear task relationship modeling: from initial linear self-representation learning to the construction of symmetric affinity matrices, and then to the dynamic generation of adaptive kernel matrices, it achieves accurate capture of complex nonlinear dependencies between tasks; by alternately optimizing and solving non-convex optimization problems, the resulting self-representation matrix Z can not only characterize pairwise affinity between tasks, but also reveal potential task groupings, providing fine-grained task relationship guidance for structure-aware aggregation. Meanwhile, the entire inference process is computationally efficient and supports complexity optimization in ultra-large-scale scenarios, ensuring good scalability in heterogeneous environments of mobile edge federated learning.

[0120] Furthermore, the structure-aware two-layer aggregation in step S5 above specifically includes the following steps:

[0121] S50. Calculation of Aggregated Weights for the Shared Encoder: The aggregated weights of the shared encoder are calculated based on the self-representation matrix Z. For each task (i,t), the cross-task influence score is defined. The affinity between the summation task (i,t) and all other task units is obtained as follows:

[0122] (27);

[0123] in, For the set of all clients, For the client The set of tasks on the platform.

[0124] Aggregate these scores by mobile edge client and globally normalize them to obtain the encoder aggregate weights:

[0125] (28);

[0126] in, For the client Local task collection on and These are the summation variables for the client index and the task index, respectively.

[0127] This weighting mechanism prioritizes edge nodes with strong structural affinity to the global task manifold, preventing low-quality updates from disrupting the shared feature extractor.

[0128] S51, Shared Encoder Update: Update the global shared encoder based on encoder aggregate weights.

[0129] (29);

[0130] S52. Calculation of Aggregated Weights for Personalized Decoder: The aggregated weights for the personalized decoder are calculated based on the self-representation matrix Z. In-task centrality scores are defined. The result is obtained by summing the bidirectional affinity between the decoder (i,t) and similar nodes aligned with its task:

[0131] (30);

[0132] in, For a set of tasks with a structure similar to task t, normalize the centrality scores to obtain the decoder aggregation weights:

[0133] (31);

[0134] in, To carry out the mission The collection of all clients.

[0135] S53, Personalized Decoder Update: Update the personalized global decoder of task t based on decoder aggregation weights.

[0136] (32);

[0137] S54, Global Model Distribution: After completing the two-layer aggregation, the edge server will update the global shared encoder. and task-specific personalized decoder sets Distribute to the corresponding mobile edge clients as the next round Enhanced initialization for local training.

[0138] In this embodiment, a differentiated aggregation strategy is designed to address the different functional roles of the shared encoder and the personalized decoder: shared encoder aggregation is based on cross-task influence scores, prioritizing the integration of client knowledge with strong structural affinity to enhance the learning of transferable features; personalized decoder aggregation is based on intra-task centrality scores, sharing knowledge only within structurally similar task clusters to preserve task specificity. This two-layer aggregation mechanism promotes positive knowledge transfer while effectively suppressing negative interference, solving the problem that traditional unified aggregation strategies cannot balance model generality and personalization, and ensuring optimal overall performance of the global model in heterogeneous task scenarios.

[0139] This invention, employing the above technical solutions, offers the following advantages compared to existing technologies: 1) The global nonlinear task relationship graph constructed through the adaptive structural reasoning module overcomes the limitations of traditional linear assumptions in complex real-world scenarios. This enables the system to more reliably identify the essential relationships between tasks in edge environments where network topology and device states dynamically change, providing accurate basis for subsequent resource allocation and model collaboration, and improving the system's decision robustness and reliability in real, unstable environments. 2) A structure-aware, two-layer aggregation strategy is adopted to implement differentiated fusion of the shared encoder and personalized decoder. This not only optimizes the utilization efficiency of limited computing and storage resources on the edge side and avoids redundant computation, but also ensures that when serving numerous heterogeneous devices, the global model possesses both the universality of common knowledge and the personalized adaptability to specific terminals or scenarios, thereby meeting the diverse Quality of Service (QoS) requirements of MEC. 3) An integrated dual semantic communication mechanism combines model and data compression with inherent privacy protection. This mechanism, through sparsification, quantization, and dedicated local codebook technologies, significantly reduces communication load and transmission latency in low-bandwidth, high-latency mobile edge networks. Meanwhile, its feature discretization and local codebook design effectively resist common edge-side data theft and gradient inversion attacks without relying on additional encryption overhead, providing feasible security technology guarantees for deploying federated learning in untrusted or open network environments. 4) Through adaptive codebook size and dynamic quantization strategies, this invention flexibly adapts to different types of edge devices, from ultra-low-power IoT sensors to high-performance edge gateways and in-vehicle mobile terminals. This invention can be efficiently deployed in real-world mobile edge computing scenarios with strictly limited resources and heterogeneous hardware, and maintains linear system scalability as the scale of edge nodes increases.

[0140] Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Without conflict, the embodiments and features in the embodiments of this application can be combined with each other. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

Claims

1. A semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing, applied in a federated learning system with one edge server and multiple mobile edge clients, characterized by: The method includes the following steps: Model distribution: The edge server builds and initializes a global model, which includes a shared encoder network for extracting cross-task shared features and multiple personalized decoder networks corresponding to the local tasks of each mobile edge client, and distributes the global model to each mobile edge client. Local training: Each mobile edge client trains the received global model based on the private heterogeneous multi-task dataset stored locally to generate local model updates. Local model updates include parameter updates for the shared encoder network and parameter updates for the personalized decoder network corresponding to the task. Parameter compression: Each mobile edge client compresses the parameter semantic representation of the local model update, and the compressed parameter semantic representation is uploaded to the edge server through the mobile edge network; The implementation of parameter semantic representation compression includes: generating and maintaining a local private codebook based on the semantic feature distribution of local multi-task data; performing error compensation and sparsification on the local model update parameters to filter out important parameter values ​​and position information; performing quantization and dequantization reconstruction on the selected important parameter values ​​to calculate residual vectors for error compensation in the next round; compressing the position index sequence of position information after index encoding to obtain a compressed index bitstream; and using the quantized residual vector to map the local task features of the current round to the local private codebook to obtain quantization indexes and update the quantized residual vectors. Relational reasoning: The edge server reconstructs local model updates based on the compressed parametric semantic representations uploaded by each mobile edge client, and infers the nonlinear global task relationship graph between tasks on all mobile edge clients through an adaptive kernel learning method, obtaining a self-representation matrix. The implementation of the adaptive kernel learning method includes: the edge server reconstructs a task representation matrix based on the compressed parametric semantic representations uploaded by each mobile edge client, and obtains an initial representation matrix through preliminary linear self-representation learning; a symmetric affinity matrix is ​​constructed based on the initial representation matrix and normalized to obtain a relationship matrix; an adaptive kernel matrix is ​​constructed based on the relationship matrix; and an alternating optimization method is used to solve the optimization problem of the adaptive kernel matrix. After convergence, the self-representation matrix is ​​learned as the global task relationship graph, and the self-representation matrix provides a structured representation of the global task manifold. Aggregation and synchronization: The edge server performs structure-aware two-layer aggregation on the shared encoder network and the personalized decoder network based on the self-representation matrix to obtain the updated global model and distribute it to each mobile edge client; and repeats the local training, parameter compression, relation inference and aggregation and synchronization steps until the global model converges and meets the end-edge collaborative training performance requirements of mobile edge computing.

2. The semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing according to claim 1, characterized in that: The implementation of generating a local private codebook includes: performing covariance analysis on the local multi-task semantic features of the mobile edge client to obtain the effective intrinsic dimension, and adaptively determining the codebook size based on the effective intrinsic dimension and the number of tasks. The local private codebook is stored locally on the client and is not uploaded.

3. The semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing according to claim 1, characterized in that: The method for compressing a compressed index bitstream after indexing and encoding the position index sequence of position information includes: sorting the parameter position index sequence, differential encoding, and applying an entropy encoding algorithm to generate a compressed index bitstream.

4. The semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing according to claim 1, characterized in that: The implementation of two-level aggregation includes: The first aggregation weight is calculated for each mobile edge client based on the self-representation matrix, and the shared encoder network parameter updates from each mobile edge client are weighted and aggregated with the first aggregation weight to update the global shared encoder network. For each task, a second aggregation weight is calculated for each mobile edge client associated with the corresponding task based on the self-representation matrix. The second aggregation weight is then used to weight and aggregate the personalized decoder network parameter updates of the relevant tasks of the associated mobile edge clients to update the global personalized decoder network of each task.

5. The semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing according to claim 4, characterized in that: The calculation of the first aggregation weight specifically includes: summing the absolute values ​​of the affinity between all tasks on each mobile edge client and tasks on other mobile edge clients in the self-representation matrix, and then globally normalizing the cumulative values ​​of each mobile edge client to obtain the weight of each mobile edge client in the shared encoder network aggregation.

6. The semantic communication privacy-enhanced federated multi-task learning method for mobile edge computing according to claim 4, characterized in that: The calculation of the second aggregation weight specifically includes: selecting mobile edge clients with affinity exceeding a preset threshold on a specified task from the self-representation matrix to form a set; calculating and normalizing the relative centrality score of each mobile edge client on the corresponding task within the set; and obtaining the weight of each mobile edge client in the personalized decoder network aggregation for the corresponding task.

Images