A Multi-Edge Collaborative Caching Method Based on Robust Federated Deep Learning

Abstract

This invention proposes a multi-edge collaborative caching method based on robust federated deep learning. First, RoCoCache designs a novel partitioning mechanism for the multi-dimensional cache space, achieving accurate content recommendation within user classification intervals. Next, RoCoCache develops a new discrete classification variational autoencoder (DC-VAE) to accurately predict content popularity by overcoming posterior collapse. Finally, RoCoCache creates a robust federated deep learning-based original training model and an active cache replacement strategy. This invention combines residual-based adversarial model update detection with similarity-based federated aggregation, avoiding model destruction caused by adversarial updates and enabling active cache replacement to adapt to optimized cache resources, thereby improving cache performance.

Description

Technical Field

[0001] This invention relates to the field of network communication technology, and in particular to a multi-edge collaborative caching method based on robust federated deep learning. Background Technology

[0002] With the significant development of 5G technology, increasingly complex applications are being deployed in fields such as industrial manufacturing, the digital economy, connected vehicles, and smart cities. Taking cloud computing as an example, during computation, tasks and data generated by applications are uploaded to remote cloud processing, causing severe network congestion and service delays. To alleviate this problem, Mobile Edge Computing (MEC), as an important technology for future networks, deploys computing and storage resources closer to the network edge, providing more reliable real-time computing and data storage for terminal device applications. Therefore, edge nodes can perform various management operations, such as signal processing, distributed caching, and wireless resource collaboration. Among these operations, distributed caching can cache user-centric content within the edge node, thereby reducing access latency and data duplication, enhancing user experience, and reducing system costs. However, caching performance is typically limited by the size and overhead of the edge cache space. Therefore, how to effectively utilize cache space and improve caching performance has received widespread attention from academia and industry. Generally, caching performance is constrained by various factors, including cache size, content relevance, cache partitioning, and cache placement strategies. Previous research has considered proactive content caching based on user preferences within a limited cache space. However, the cache space can be further partitioned based on user characteristics, activity, and request patterns. This approach optimizes the configuration of edge cache resources, enabling accurate content recommendations for similar users across multiple edge nodes. Exploring the potential relationships between user and content characteristics in a multi-dimensional space can improve the hit rate of user access to resources, but it remains an open challenge.

[0003] Multi-edge collaborative caching has emerged as a viable approach to optimize cache resource configuration and reduce service latency. If a user's connected edge node doesn't match their request, the user can determine the requested content from other edge nodes performing collaborative caching. However, existing research extensively employs deep reinforcement learning (DRL) and game theory methods. With the rapid increase in the number of terminal devices and growing sensitivity to user privacy, these methods may no longer be optimal for addressing security concerns and suboptimal cache resource configuration challenges. Federated learning (FL), as a distributed training framework, is considered a promising solution to these problems. Following its core principle, edge nodes collaboratively train a global model by uploading model parameters without revealing raw data. FL maximizes the satisfaction of heterogeneous user needs across multiple edge nodes through a well-trained global model, while ensuring user privacy and the scalability of large-scale network models. However, in complex MEC environments, unintentional model corruption or malicious adversarial model interference can lead to the inability to perform model training and degraded model update quality. For example, unintentional model corruption can occur due to noisy training labels, insufficient data samples, and low-quality uploaded models. Furthermore, malicious edge nodes may intentionally launch adversarial attacks to tamper with the model, such as Byzantine attacks and backdoor attacks. In addition, detecting adversarial attacks requires significant computation, as it necessitates substantial data collection and analysis to identify improper behavior. Research indicates that running a previous intrusion detection system (IDS) on a 10MB dataset consumes over 400 seconds of CPU time, 200 joules of energy, and 100MB of RAM on mobile devices, severely impacting its normal operation. These issues not only exacerbate security challenges but also degrade the performance of the global model, affecting cache hit rates and increasing system overhead. Therefore, applying Robust Federated Deep Learning (RFDL) to edge caching is crucial for enhancing system security and optimizing cache performance.

[0004] Specifically, the following summarizes the main challenges of applying FL to handle multi-edge collaborative caching issues:

[0005] 1) Discrete user characteristic distribution and the diversity of requested content: Different users may have different content preferences due to their discrete nature. Therefore, finding the potential relationship between discrete user characteristic distribution and diverse content requests is a challenge.

[0006] 2) Model Scalability: As the number of terminal devices increases, more data will be discretely distributed, leading to higher computational and communication overhead in caching scenarios with a single edge node. However, classic centralized training frameworks encounter limited model scalability.

[0007] 3) Model robustness: Edge nodes may unintentionally upload low-quality models or be subjected to adversarial attacks by malicious nodes, resulting in a severe decrease in robustness during model updates.

[0008] To address the aforementioned challenges, this invention proposes a novel collaborative caching framework based on robust federated deep learning (RoCoCache). First, RoCoCache designs a novel partitioning mechanism for the multidimensional cache space, achieving accurate content recommendation within user classification intervals. Next, RoCoCache develops a novel discrete classification variational autoencoder (DC-VAE) to accurately predict content popularity by overcoming posterior collapse. Finally, RoCoCache creates a robust federated deep learning-based original training model and an active cache replacement strategy. Notably, combining residual-based adversarial model update detection with similarity-based federated aggregation avoids model corruption caused by adversarial updates, enabling active cache replacement to adapt to optimized cache resources, thereby improving cache performance. Summary of the Invention

[0009] This invention proposes a multi-edge collaborative caching method based on robust federated deep learning, which can improve caching performance.

[0010] The present invention adopts the following technical solution.

[0011] A robust federated deep learning-based multi-edge collaborative caching method is proposed for a multi-edge collaborative caching system consisting of M edge nodes, denoted by a set E = {e1, e2, ..., e...}. m , ..., e M} indicates that each node contains one MEC server, one base station, and N users, represented by the set U = {u1, u2, ..., u...}. n , ..., u N} represents the cache space of an edge node, defined as a set C = {C1, C2, ..., C}. m C M Each user connects to an edge node and communicates via a wireless link provided by the associated base station; furthermore, communication between edge nodes and between edge nodes and the cloud data center is conducted via a backhaul link; the cache space status of each edge node is periodically broadcast to other edge nodes within the proposed system; the content library of the cloud data center is represented as F = {f1, f2, ..., f...} i , ..., f I}, where I represents the number of accessible contents; and users are discretely distributed across the service areas of various edge nodes;

[0012] When user u n Send content f to the edge nodes it is connected to i When making a request, the multi-edge collaborative caching method RoCoCache (Collaborative Caching framework based on Robustfederated deep learning, RoCoCache) for multi-edge collaborative caching scenarios includes the following steps;

[0013] Step S1: The current edge node checks whether f is cached. i If f i If cached, the edge node will send it directly to u. n Otherwise, proceed to step S2;

[0014] Step S2: Search the current edge node for the existence of cache f. i The collaborative edge node. If it exists, the collaborative edge node will transmit f via the backhaul link. i Forward to the current edge node, then f i Send to u n Otherwise, proceed to step S3;

[0015] Step S3: If there is no cooperating edge node cache f i The content library of the cloud data center will provide f i And forward it to the current edge node via the backhaul link, and then f i Send to u n .

[0016] In the multi-edge collaborative caching scenario, content requests from users are dynamic and have spatiotemporal dependencies; improving cache hit rate relies on accurate content popularity prediction and caching content that users are interested in into the cache space of edge nodes; specifically, f i At edge node e m Popularity on the platform is defined as

[0017]

[0018] where req i，m For e m Received f i Request count, req m For e m Total number of requests received;

[0019] The RoCoCache method accurately predicts the popularity of content; to evaluate prediction accuracy, the global loss function is defined as follows:

[0020]

[0021] In the formula, r represents the FL communication wheel, and w (r) Here are the parameters for the global prediction model, and req is the total number of requests received by all edge nodes. For the model parameters of the local FL, the mean squared error (MSE) is defined as follows:

[0022]

[0023] in f i For e m The predicted popularity value, P i，m (r) represents the actual value;

[0024] Furthermore, the cache hit rate is defined as...

[0025]

[0026] Where θ m (f i ) represents e m Whether to cache the content requested by the user is defined as follows:

[0027]

[0028] Cache performance is affected by various factors, including cache resource configuration, content popularity, model robustness, and cache replacement strategy. By comprehensively considering these factors, the RoCoCache multi-edge collaborative caching method can effectively improve the cache performance of multi-edge collaborative caching systems. The RoCoCache method can be used in the following ways:

[0029] A. Through a multi-dimensional edge cache space partitioning mechanism, cache resources are optimized to ensure that users receive accurate content recommendations within their category range. This mechanism includes multi-dimensional user partitioning and cache space partitioning, taking into account differences in user characteristics, activities, and memory access ranges.

[0030] B. The RoCoCache method is based on the developed Discrete Classification Variational Autoencoder (DC-VAE) to accurately adapt to the discrete distribution in the user request matrix. DC-VAE first learns discrete vectors and then uses nearest neighbors to find discrete latent vectors to assist the decoder in generating the user request matrix.

[0031] The C and RoCoCache methods are trained using user request data stored on each edge node, and aggregate local models to generate a globally shared model. Specifically, a residual-based detection method is proposed to accurately capture adversarial model updates, and a similarity-based FL aggregation method is used to avoid adversarial updates from damaging the globally shared model.

[0032] The multi-edge collaborative caching method RoCoCache adopts the multi-edge system elastic collaborative caching framework RoCoCache based on robust federated deep learning RFDL. Specifically, RoCoCache optimizes the cache space of edge nodes by partitioning the cache space in a multi-dimensional manner and determines an appropriate cache space size for users in the interval.

[0033] Secondly, RoCoCache uses DC-VAE to learn the latent space containing discrete vectors, forming a content popularity prediction based on DC-VAE. In DC-VAE, the decoder uses the nearest neighbor algorithm to find discrete latent vectors, and then generates a calibrated user request matrix to improve the accuracy of content popularity prediction.

[0034] Then, RoCoCache uses an RFDL-based training mode to improve the scalability and robustness of the model; the training mode includes a residual-based detection method to capture adversarial model updates, and also includes a similarity-based FL aggregation method to avoid the damage caused by adversarial updates to the globally shared model;

[0035] Finally, an active cache replacement strategy based on RFDL is adopted to adapt to the optimized cache resource configuration and improve the performance of multi-edge collaborative caching.

[0036] The multidimensional cache space partitioning includes two key components: multidimensional user partitioning and cache space partitioning. In multidimensional user partitioning, RoCoCache classifies and segments feature groups with different numbers of users, and caches the content that users are interested in for different groups. In cache space partitioning, the cache space is perceptually optimized based on user characteristics, user activity, and the dispersion of memory access intervals, using the established classification.

[0037] The multidimensional user segmentation specifically refers to: assuming that user characteristics reflect their preferences for cached content to a certain extent; such as... Figure 1 As shown, to accurately predict user preferences, RoCoCache uses a user segmentation method based on multidimensional features, including gender, age, and occupation. These features are continuously encoded as coordinate axes, using a set Γ = {γ1, γ2, ..., γ...}. t , ..., γ TThe representation} makes it possible to construct a feature hypercube, in which users are grouped to form user intervals; as the partitioning proceeds, the user intervals are represented as the set H = {h1, h2, ..., h...}. s , ..., h S} where S is the number of user intervals, Grade(h s This represents the user hierarchy within a user region; the side length of each hypercube is represented as a set. In the initial stage (Grade(h0) = 0), all users with different characteristics are placed in the same user interval h0; if the number of users in |h0| exceeds the threshold ζ(Grade(h0)), they are divided into 2 equal parts along each dimension. T There are several intervals, where the length of each dividing dimension will be halved, i.e. ζ(Grade(h s The value of ζ(Grade(h)) determines the number of users within the user range; when ζ(Grade(h) s A larger ζ(Grade(h)) results in a higher user density within each interval, thus affecting the ability to accurately capture different user preferences; conversely, a smaller ζ(Grade(h)) s The value corresponds to a reduced user density, which may lead to inaccurate cache predictions.

[0038] To achieve adaptive segmentation of user regions and capture the potential relationship between users within a region and their preferred content, set Where α is a hyperparameter. The division of each user interval may continue and proceed to the following stages based on performance requirements, with the stage representation including Grade = 1, 2, ...

[0039] The cache space allocation improves cache performance by considering various factors, including the number of users, user activity, and the dispersion of memory access ranges, and allocating appropriate cache space to user ranges. Specifically:

[0040] In the divided user interval h s In this context, the number of users is denoted as |h s The divergence between user activity and memory access range is defined as...

[0041]

[0042] In the formula req s Number of user requests for h s The memory access range.

[0043] Taking the above factors into account, it will be assigned to h. s The cache size is defined as

[0044]

[0045] In the formula, For the cumulative product function, cache m For connecting h s The size of the cache space on the edge nodes.

[0046] The content popularity prediction based on DC-VAE is as follows: Assume that the popularity of content in real-world scenarios typically follows a power-law distribution and always adheres to Zipf's law. In content popularity prediction, a Variational Autoencoder (VAE) is used to efficiently reconstruct the compressed input data using continuous variables in the hidden layer, enabling data clustering within the latent space. Specifically, the VAE consists of three main parts: an encoder, a latent space, and a decoder. Given input data x, the encoder output parameters are μ and log(σ). 2 Gaussian distribution Where z is the latent representation; next, the VAE sample z is derived from... The parameterization is redefined in the middle, where z = μ + σ ∈; the decoder is denoted as p. θ (x|z), reconstructing the compressed input data x; in this process, the reconstruction error is minimized using the evidence lower bound ELBO and the distribution of the latent space is regularized, defined as

[0047]

[0048] The first term is used for reconstruction, and KL(·) represents the (Kullback-Leibler KL) divergence KL used for regularization; these two terms are used to achieve a balance between fitting the posterior and maintaining the latent space distribution; considering the substantial impact of significant noise on the latent space representation in the VAE, and this impact makes μ and log(σ) equal. 2 The estimated value of ) may become unstable, thus presenting an elusive latent variable z; in this case, the regularization term in formula (9) loses its effectiveness; it cannot enforce the prior distribution; then the decoder abandons the use of the latent space, resulting in a generalized output x, characterized by posterior collapse; when facing weak signals, the approximate posterior distribution q φ (z|x) tends to approximate the prior distribution q φ (z), using This indicates that the KL divergence is invalid. In order to prevent the VAE from being trapped in the local optimum space, the Discrete Classification Variational Autoencoder (DC-VAE) is used. Learnable discrete vectors are used to form the latent category space. When predicting content popularity, the goal of DC-VAE is to find the vector with the closest encoding distance to the encoder network output in the latent category space, and then reconstruct the mapped vector through the decoder network.

[0049] Figure 2 This demonstrates content popularity prediction based on DC-VAE; specifically, DC-VAE learns the latent distribution in the user request matrix X with the aim of obtaining future user requests in the reconstruction matrix output by the decoder; the user request matrix X contains historical information about user request content at edge nodes, defined as...

[0050]

[0051] In the formula, l≤n≤N, and n represents the number of users connected to a certain edge node; This represents the content request record for user n, where l ≤ i ≤ I, and i is the content library index; This indicates that the content request was successful; This indicates either a content request failure or a lack of interest in the content, but these two situations are difficult to distinguish, leading to inaccurate predictions. To address this issue, matrix X was supplemented and calibrated; specifically, in DC-VAE, the latent space is defined as... Where K is the size of the space and D is the dimension of the vector; therefore, there are K category vectors. like Figure 2 As shown, the DC-VAE inputs x through the encoder network. n Output t v (x n );

[0052] Next, the discrete latent variable t and the posterior probability distribution q(t|x) are calculated using the nearest neighbor algorithm. n ) is a one-hot encoding, defined as

[0053] The input to the decoder is defined as t q (x n ) = v k Where k is the index of the decoder input, defined as k = argmin j ||t v (x n →v j ||2;

[0054] To address the gradient collapse problem caused by the introduction of implicit spaces, the gradient V is transformed during backpropagation. z L is copied from the decoder network to the encoder network; during training of the DC-VAE, the loss function is defined as...

[0055]

[0056] Where log p(x) n |t q (xn The reconstruction loss is used to optimize the encoder and decoder networks; since the backpropagation gradient is directly copied to the encoder network, the loss log p(t|t) is not considered. q (x n ));

[0057] exist In the middle, the L2 error is used to measure v k Drive to t v (x n This aims to optimize the implicit category space; This is to prevent the encoder output from exceeding the range of the latent class space, where λ depends on the reconstruction loss, and sg is the stopping gradient operator, which is a constant during forward propagation and has a partial derivative of 0.

[0058] Next, the log-likelihood function is defined as follows:

[0059] log p(x n )≈log p(x n |t q (x n ))p(t q (x n (13)

[0060] According to Jensen's inequality, equation (13) can be rewritten as follows:

[0061] log p(x n )≥log p(x n |t q (x n ))p(t q (x n (14).

[0062] The robust federated deep learning RFDL includes residual-based detection and similarity-based federated aggregation. Residual-based detection detects adversarial model updates by ranking parameters. Similarity-based federated aggregation aims to avoid adversarial updates from damaging the globally shared model and generate robust and accurate content popularity prediction models in complex multi-edge environments.

[0063] The residual-based detection specifically involves designing a parameter sorting matrix to prevent adversarial model updates from affecting the model's robustness. To detect adversarial updates; such as Figure 3 As shown, the model parameters from all edge nodes are first combined into a matrix. Defined as

[0064]

[0065] Among them, the global sharing model uses w (*) Parameterization, d m This represents local training data;

[0066] Next, sort the elements in each column of R in descending order, preserving their sort order, and convert them to... Will The mean and variance var are defined as follows:

[0067]

[0068] Based on the mean and var, normal and adversarial model updates can be clustered into two groups using K-means, where adversarial model updates can be easily identified using the proposed residual-based detection method. By using the proposed residual-based detection method, edge nodes providing normal model updates can be filtered, denoted as E. b ={e b，1 e b，2 , ..., e b，M′}

[0069] The similarity-based federated aggregation employs canonical correlation analysis (CCA) to measure the similarity between the model update of each edge node and the average, thereby determining the weights of different model updates when performing federated aggregation; described as follows:

[0070]

[0071] Among them κ m This indicates the similarity score;

[0072] The RFDL algorithm combines residual-based detection with similarity-based federated aggregation, and includes the following steps;

[0073] Cloud data center update; First, set the total number r of FL communication wheels. max And initialize the global content popularity prediction model w (r) For each FL communication round, edge nodes update their local models in parallel; next, residual-based detection identifies adversarial model updates and obtains the edge node E that provides the normal model update. b Finally, a globally shared model is generated through similarity-based federated aggregation and distributed to edge nodes;

[0074] Update at each edge node; first, initialize the training epoch c. max Batch size B and learning rate η; with a globally shared model w (r)The input is used by each edge node to begin its own local training; for each training epoch, DC-VAE uses the Adam optimizer in batches to train and update the local model; after local training, each edge node uploads its latest local model to the cloud data center; D. Active cache replacement with RFDL

[0075] Based on the proposed RFDL, an active cache replacement strategy with multi-edge collaboration is designed; the key steps are as follows: For each edge node, the cache space is initialized through multi-dimensional cache space partitioning. temp And the user interval set H, when cache temp When the value is ≥0, RFDL is invoked to predict and rank content popularity, and content that the user is interested in will be placed in a temporary cache library C. temp In order to avoid cache redundancy caused by overlapping user-interested content in different intervals, C will be used. temp Replace with C s From C temp Select the current user range h s The most popular cache h Content; Next, delete the cache library C on each edge node. m The process involves identifying duplicate entries in the cache and updating the available cache space; this process is repeated until the cache space is fully utilized.

[0076] This invention proposes a novel collaborative caching framework based on robust federated deep learning (RoCoCache). First, RoCoCache designs a novel partitioning mechanism for the multi-dimensional cache space, achieving accurate content recommendation within user classification intervals. Next, RoCoCache develops a new discrete classification variational autoencoder (DC-VAE) to accurately predict content popularity by overcoming posterior collapse. Finally, RoCoCache creates a robust federated deep learning-based original training model and an active cache replacement strategy. Notably, combining residual-based adversarial model update detection with similarity-based federated aggregation avoids model corruption caused by adversarial updates, enabling active cache replacement to adapt to optimized cache resources, thereby improving cache performance.

[0077] This invention designs RoCoCache, a resilient collaborative caching framework that uniquely integrates RFDL with an active caching strategy. First, a multi-dimensional cache space partitioning mechanism is designed to optimize the edge cache space, providing accurate content recommendations within user interval categories. Next, a content popularity prediction algorithm based on DC-VAE is developed, solving the posterior collapse problem and improving prediction accuracy. Finally, an RFDL-based training mode and an active cache replacement strategy are established to improve adaptability and robustness in complex network environments. Using real-world testing platforms and datasets, RoCoCache achieves a higher cache hit rate than state-of-the-art methods and approaches the best hit rate. Ablation experiments verify that the multi-dimensional cache space partitioning and collaborative caching design can improve caching performance. Furthermore, RoCoCache exhibits excellent training and caching efficiency across various scenarios. Additionally, RoCoCache can identify adversarial model updates in complex network environments, validating its good robustness. Attached Figure Description

[0078] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0079] Appendix Figure 1 This is a schematic diagram of the multidimensional cache space partitioning in this invention;

[0080] Appendix Figure 2 This is a schematic diagram of the DC-VAE content popularity prediction model in this invention;

[0081] Appendix Figure 3 This is a schematic diagram of the RFDL model in this invention;

[0082] Appendix Figure 4 This is a schematic diagram comparing the implementation with the State-of-the-Arts method;

[0083] Appendix Figure 5 This is a comparative schematic diagram of the ablation experiment in the embodiments;

[0084] Appendix Figure 6 This is a comparative schematic diagram of convergence analysis in the embodiments;

[0085] Appendix Figure 7 This is a schematic diagram comparing training efficiency in the embodiments;

[0086] Appendix Figure 8 This is a comparative schematic diagram of model detection in the embodiment;

[0087] Appendix Figure 9 This is a comparative schematic diagram of robust performance in the embodiments; Detailed Implementation

[0088] In this example, the proposed RoCoCache is evaluated through comparative experiments.

[0089] A. Experimental Setup

[0090] Real-world testbed. A real-world testbed was built, consisting of one workstation and 20 Jetson TX2s. The workstation, serving as a cloud data center, is equipped with two NVIDIA GeForce GTX 3090 GPUs, one Intel(R) Xeon(R) CPU Silver4208 @ 2.10GHz, and 32GB of RAM. The Jetson TX2s act as edge nodes, each equipped with an NVIDIA Pascal GPU with 256 CUDA cores and a CPU processor consisting of two Denver2 cores and four ARM Cortex A57 cores. The workstation and Jetson TX2s are deployed on the same network. A backend service based on the FLASK web framework was built to facilitate communication between the workstation and the Jetson TX2s. The testbed equipment runs on Ubuntu 18.04, CUDA v10.0, and cuDNN v7.5.0.

[0091] Datasets. Three real-world datasets were used: MovieLens 100K and 1M, and iQiyi's dataset. These three datasets contain approximately 100,000, 1,000,000, and 10,000,000 request records, respectively. The MovieLens 100K and 1M datasets were collected by groupLens Research. MovieLens 100K contains approximately 10,000 ratings from 943 anonymous users on 1,682 movie projects, while MovieLens 1M contains approximately 1,000,000 ratings from 6,040 anonymous users on 3,883 movie projects. Additionally, the iQiyi dataset released by iQiyi contains approximately 10 million requests from 100,000 users on 1,900,000 video projects. These datasets provide user sequence numbers, project indices, timestamp labels, and user context information. Specifically, user gender, age, and occupation were selected as user characteristics, and ratings were considered user requests. The dataset is divided into training set (70%), validation set (10%), and test set (20%).

[0092] Parameter settings. Based on the aforementioned real-world testbed and dataset, a multi-edge collaborative caching scenario was simulated, consisting of a cloud data center, 5–20 edge nodes, and 9.43–100,000 users. The cloud data center stores the complete real-world dataset, each edge node is equipped with a fixed-size cache space, and users are randomly distributed across the service areas of each edge node. RoCoCache was implemented using Python 3.8 and Tensorflow 2.4.0. Specifically, the hyperparameter α in the multidimensional cache space partitioning is 512, the size of the DC-VAE hidden class space K is 128, and the dimension Dv of the class vectors is... c The number of FL communication rounds is 16, r. max The batch size in DC-VAE is 32, and the number of training epochs is c. max The value is 300, and the learning rate η is 0.001. According to experimental tests, the latency for accessing content requests to local edge nodes, collaborative edge nodes, and cloud data centers is approximately [2, 4] ms, [15, 20] ms, and [18, 23] ms, respectively.

[0093] Comparison methods. RoCoCache is compared with the best and the following state-of-the-art methods.

[0094] (1) Oracle: It can predict all prior information about future user requests, so it can achieve the best cache hit rate with limited cache space.

[0095] (2) Random: Randomly select the content requested by the user for active caching.

[0096] (3) LRU (Least Recently Used): Sorts the least recently used content based on the request time.

[0097] (4) Learning-based collaborative caching strategy (LECS): Combines a content popularity prediction model based on temporal convolutional networks with a dynamic programming algorithm for collaborative content caching.

[0098] (5) Edge Collaboration Cache (ECC): Integrates a neural collaboration filter for content popularity prediction and a greedy algorithm for content delivery.

[0099] (6) Autoencoder (AE): First, the input data is reconstructed by compressing the hidden layer of the encoder, and then the predicted content popularity distribution is obtained from the output matrix.

[0100] (7) Variational Autoencoder (VAE): Improved AE and uses continuous variables in the hidden layer to reconstruct compressed input data.

[0101] Attack Models. The robustness of RoCoCache was evaluated using the following two attack models.

[0102] Symbol Flip Attack (SFA): Generates adversarial model updates by reversing normal model updates, denoted as... Where μ > 0.

[0103] Gaussian Noise Attack (GNA): Generates adversarial model updates by adding Gaussian random noise to normal model updates.

[0104] B. Experimental Results and Analysis

[0105] Comparison with State-of-the-Arts methods. Comparative experiments on cache hit rates were conducted on various datasets with different edge cache space sizes. For example... Figure 4 As shown, the cache hit rate of all methods increases with the increase of the edge cache space size. Oracle achieves theoretically optimal results because it anticipates prior information about future user requests. Random caching shows the worst performance because its caching strategy is blind. AE and VAE exhibit good cache hit rates because they compress high-dimensional user requests into low-dimensional representations and learn the potential relationship between user features and requested content. However, due to the difference between large data volumes and the assumption of a single distribution, AE suffers from posterior collapse on the iQiyi dataset. Therefore, AE shows an extremely low cache hit rate. By using clustering in the continuous latent space, VAE has a better ability to reconstruct the input distribution than AE, thus achieving more accurate content popularity predictions and a higher cache hit rate. Furthermore, ECC shows a decreasing trend in cache hit rate with increasing edge cache space, consistently performing worse than RoCoCache. This is because the collaborative filter used by ECC cannot accurately capture content popularity, resulting in inefficient utilization of the edge cache space. LECS employs a content popularity prediction model based on a temporal convolutional network, achieving a growth rate comparable to other popularity-driven methods. However, when content popularity exhibits similar patterns but differs in magnitude of change, this temporal neural network struggles to establish clear boundaries for effective caching, ultimately reducing cache hit rate. LRU responds well to bursty and sparse content requests but struggles to handle changing trends in content popularity. Therefore, LRU's caching performance is inferior to AE, VAE, ECC, and LECS. Compared to other methods, RoCoCache demonstrates a higher cache hit rate, approaching the optimal level. This is because RoCoCache achieves perceptual optimization of the edge cache space through multi-dimensional cache space partitioning, while simultaneously addressing the posterior collapse problem in VAE, resulting in more accurate content popularity prediction.

[0106] Ablation experiments were conducted on the MovieLens 1M dataset to evaluate the effectiveness of the multidimensional cache space partitioning and cooperative caching design in RoCoCache. Figure 5 As shown, the cache hit rate of all methods increases with the increase of the edge cache space size. Random caching shows the worst performance because its caching strategy is blind. Since DC-VAE can achieve near-optimal results without cooperative caching, its performance is comparable to RoCoCache, which only uses multidimensional cache space partitioning. When RoCoCache adopts cooperative caching, it can benefit from the joint optimization of multiple edge cache resources, thus achieving a higher cache hit rate. It is worth noting that RoCoCache achieved the best cache hit rate in all ablation experiments, indicating that multidimensional cache space partitioning can effectively assemble the common user preferences of each edge node. Therefore, RoCoCache maintains good caching performance in both cooperative and non-cooperative caching scenarios. The cache hit rate of RoCoCache under attacks is closely consistent with the overall hit rate of RoCoCache and will be thoroughly analyzed in the robustness analysis section.

[0107] Convergence analysis. Figure 6 This paper analyzes the convergence of RoCoCache on the MovieLens 1M dataset as the number of FL communication rounds increases. Different sizes of edge cache spaces were considered in the experiments to ensure comprehensive testing. In the initial stage, content was randomly selected and stored in the cache spaces of edge nodes, resulting in a low cache hit rate. After one round of FL communication, RoCoCache generates a preliminary prediction model of content popularity, thereby rapidly improving the cache hit rate. At this point, under different edge cache space sizes, RoCoCache can achieve over 80% of the optimal caching performance, where the size of the edge cache space determines the range of cache hit rate growth. Notably, in different scenarios, RoCoCache shows a convergence trend after 6 rounds of FL communication, demonstrating its excellent convergence performance.

[0108] Training efficiency. The training efficiency of RoCoCache on the MovieLens 1M dataset was evaluated under scenarios with varying numbers of edge nodes. Figure 7As shown, the training time for each round of FL decreases as the number of edge nodes increases. With a fixed number of user content requests, adding edge nodes for collaborative caching can effectively improve the training efficiency of RoCoCache. Furthermore, with an increase in the number of edge nodes, RoCoCache can better capture different user preferences and improve cache hit rate through multi-edge collaboration. Experimental results demonstrate that RoCoCache can adapt to various multi-edge scenarios while exhibiting excellent training efficiency and cache hit rate.

[0109] Table 5-1 Comparison of content request latency (ms) between RoCocache and other methods

[0110]

[0111] Caching efficiency. The caching efficiency of different methods was tested across 10 edge node scenarios, focusing on content request latency. Non-cooperative caching (RoCoCache without collaborative caching) and distributed caching (caching only one copy of the content on each edge node based on its popularity) were evaluated. As shown in Table 5-1, content request latency decreased with increasing edge cache size. RoCoCache handles content requests in different ways and accurately predicts content popularity, thus exhibiting the best caching efficiency. Non-cooperative caching, lacking collaborative caching, requires forwarding requests for missing content to the remote cloud, resulting in lower caching efficiency. Furthermore, due to the low cache hit rate, distributed caching requires continuously sending content requests to other edge nodes and the remote cloud. Therefore, compared to RoCoCache, the other two methods lead to excessively high latency.

[0112] Robustness analysis. The robustness of RoCoCache was evaluated from two aspects. First, RoCoCache's ability to detect adversarial model updates was tested. Figure 8 The performance of the residual-based detection model is shown when exposed to 30% and 40% of model updates as adversarial. The separation between adversarial and normal model updates indicates the difference in mean and variance of the gradient parameters between the two update types. For detecting SFA, the increased proportion of adversarial model updates makes it more difficult to use residual-based detection. In this case, RoCoCache can still distinguish adversarial model updates. For GNA detection, a larger proportion of Gaussian noise significantly affects the variance of the parameter ordering matrix, making the separation more pronounced.

[0113] Secondly Figure 9This paper describes the robustness performance of RoCoCache under different attacks and defenses when the adversarial model update ratio is 30%. Without attacks, RoCoCache achieves an ideal cache hit rate. Under SFA and GNA attacks, RoCoCache still converges after approximately 20 rounds of FL communication, approaching the ideal result. Without defenses, SFA and GNA severely impact cache performance. In this case, it is clear that adjusting the cache strategy is unlikely to significantly improve the cache hit rate. SFA leads to sign inversion, severe joint aggregation, and invalidation of the globally shared model. Meanwhile, without defenses, GNA severely impacts cache performance. This is because Gaussian noise alters the weighted mean and geometric median of the globally shared model, increasing the difficulty of model training. The results validate the strong robustness of RoCoCache, provide accurate identification of adversarial model updates in complex network environments, and ensure rapid model convergence.

Claims

1. A multi-edge collaborative caching method based on robust federated deep learning, used in multi-edge collaborative caching systems, characterized by: The multi-edge collaborative caching system consists of M edge nodes, denoted by the set E = {e1, e2, ..., e...}. m , ..., e M } indicates that each node contains one MEC server, one base station, and N users, represented by the set U = {u1, u2, ..., u...}. n , ..., u N } represents the cache space of an edge node, defined as a set C = {C1, C2, ..., C}. m C M Each user connects to an edge node and communicates via a wireless link provided by the associated base station; furthermore, communication between edge nodes and between edge nodes and the cloud data center is conducted via a backhaul link; the cache space status of each edge node is periodically broadcast to other edge nodes within the system; the content library of the cloud data center is represented as F = {f1, f2, ..., f...} i , ..., f I }, where I represents the number of accessible contents; and users are discretely distributed across the service areas of various edge nodes; When user u n Send content f to the edge nodes it is connected to i When making a request, the multi-edge collaborative caching method RoCoCache is used in multi-edge collaborative caching scenarios. Includes the following steps; Step S1: The current edge node checks whether f is cached. i If f i If cached, the edge node will send it directly to u. n Otherwise, proceed to step S2; Step S2: Search the current edge node for the existence of cache f. i The collaborative edge node; if it exists, the collaborative edge node will transmit f via the backhaul link. i Forward to the current edge node, then f i Send to u n Otherwise, proceed to step S3; Step S3: If there is no cooperating edge node cache f i The content library of the cloud data center will provide f i And forward it to the current edge node via the backhaul link, and then f i Send to u n ; The multi-edge collaborative caching method RoCoCache adopts the multi-edge system elastic collaborative caching framework RoCoCache based on robust federated deep learning RFDL. Specifically, RoCoCache optimizes the cache space of edge nodes by partitioning the cache space in a multi-dimensional manner and determines an appropriate cache space size for users in the interval. Secondly, RoCoCache uses DC-VAE to learn the latent space containing discrete vectors, forming a content popularity prediction based on DC-VAE. In DC-VAE, the decoder uses the nearest neighbor algorithm to find discrete latent vectors, and then generates a calibrated user request matrix to improve the accuracy of content popularity prediction. Then, RoCoCache uses an RFDL-based training mode to improve the scalability and robustness of the model; the training mode includes a residual-based detection method to capture adversarial model updates, and also includes a similarity-based FL aggregation method to avoid the damage caused by adversarial updates to the globally shared model; Finally, an active cache replacement strategy based on RFDL is adopted to adapt to the optimized cache resource configuration and improve the performance of multi-edge collaborative caching.

2. The multi-edge collaborative caching method based on robust federated deep learning according to claim 1, characterized in that: In the multi-edge collaborative caching scenario, content requests from users are dynamic and have spatiotemporal dependencies. Improving cache hit rate relies on accurate content popularity prediction and caching content that users are interested in into the cache space of edge nodes; Specifically, f i At edge node e m Popularity on the platform is defined as where req i，m For e m Received f i Request count, req m For e m Total number of requests received; The RoCoCache method accurately predicts the popularity of content; to evaluate prediction accuracy, the global loss function is defined as follows: In the formula, r represents the FL communication wheel, and w (r) Here are the parameters for the global prediction model, and req is the total number of requests received by all edge nodes. For the model parameters of the local FL, the mean squared error (MSE) is defined as follows: in f i For e m The predicted popularity value, P i，m (r) represents the actual value; Furthermore, the cache hit rate is defined as... Where θ m (f i ) represents e m Whether to cache the content requested by the user is defined as follows: Cache performance is affected by a variety of factors, including cache resource configuration, content popularity, model robustness, and cache replacement strategy. By comprehensively considering these factors, the multi-edge collaborative caching method RoCoCache can effectively improve the cache performance of multi-edge collaborative caching systems.

3. The multi-edge collaborative caching method based on robust federated deep learning according to claim 1, characterized in that: The RoCoCache multi-edge collaborative caching method can be used in the following ways: Method A: Through a multi-dimensional edge cache space partitioning mechanism, cache resources are perceived and optimized to ensure that users receive accurate content recommendations within their category range. This mechanism includes multi-dimensional user partitioning and cache space partitioning, taking into account differences in user characteristics, activities, and memory access ranges. Method B, RoCoCache, is based on the developed Discrete Classification Variational Autoencoder (DC-VAE) to accurately adapt to the discrete distribution in the user request matrix. DC-VAE first learns from discrete vectors, and then uses nearest neighbors to find discrete latent vectors to assist the decoder in generating the user request matrix. The C-method and RoCoCache method use user request data stored on each edge node for training, and aggregate local models to generate a global shared model. Specifically, a residual-based detection method is proposed to accurately capture adversarial model updates, and a similarity-based FL aggregation method is used to avoid adversarial updates from damaging the global shared model.

4. The multi-edge collaborative caching method based on robust federated deep learning according to claim 1, characterized in that: The multidimensional cache space partitioning includes two key components: multidimensional user partitioning and cache space partitioning. In multidimensional user partitioning, RoCoCache classifies and segments feature groups with different numbers of users, and caches the content that users are interested in for different groups. In cache space partitioning, the established classification is used to perform perceptual optimization of the cache space based on user characteristics, user activities, and the dispersion of memory access intervals.

5. The multi-edge collaborative caching method based on robust federated deep learning according to claim 4, characterized in that: The multidimensional user segmentation specifically involves: assuming that user characteristics reflect their preferences for cached content to a certain extent; to accurately predict user preferences, RoCoCache uses a user segmentation method based on multidimensional features, including gender, age, and occupation. These features are continuously encoded as coordinate axes, using a set Γ = {γ1, γ2, ..., γ...}. t , ..., γ T The representation} makes it possible to construct a feature hypercube, in which users are grouped to form user intervals; as the partitioning proceeds, the user intervals are represented as the set H = {h1, h2, ..., h...}. s , ..., h S }, where S is the number of user intervals, Grade(h s This represents the user hierarchy within a user region; the side length of each hypercube is represented as a set. In the initial stage, Grade(h0) = 0, and all users with different characteristics are placed in the same user interval h0; let |h0| represent the number of users in h0. If the number exceeds the threshold ζ(Grade(h0)), then it is divided into 2 equal parts along each dimension. T There are several intervals, where the length of each dividing dimension will be halved, i.e. ζ(Grade(h s The value of ζ(Grade(h)) determines the number of users within the user range; when ζ(Grade(h) s A larger ζ(Grade(h)) results in a higher user density within each interval, thus affecting the ability to accurately capture different user preferences; conversely, a smaller ζ(Grade(h)) s The value corresponds to a reduced user density, which may lead to inaccurate cache predictions. To achieve adaptive segmentation of user regions and capture the potential relationship between users within a region and their preferred content, set Where α is a hyperparameter; according to performance requirements, it enters the following stages, and the stage description includes Grade = 1, 2, ...

6. The multi-edge collaborative caching method based on robust federated deep learning according to claim 4, characterized in that: The cache space allocation improves cache performance by considering various factors, including the number of users, user activity, and the dispersion of memory access ranges, and allocating appropriate cache space to user ranges. Specifically: In the divided user interval h s In this context, the number of users is denoted as |h s |; Define the divergence between user activity and memory access range as In the formula req s Number of user requests for h s The memory access range; Taking the above factors into account, it will be assigned to h. s The cache size is defined as In the formula, For accumulation functions, cache m For connecting h s The size of the cache space on the edge nodes.

7. The multi-edge collaborative caching method based on robust federated deep learning according to claim 4, characterized in that: The content popularity prediction based on DC-VAE specifically involves: assuming that the popularity of content in real-world scenarios typically follows a power-law distribution and always adheres to Zipf's law; using a variational autoencoder (VAE) in content popularity prediction, the continuous variables in the hidden layer are used to efficiently reconstruct the compressed input data, enabling data clustering within the latent space; specifically, the VAE consists of three main parts: an encoder, a latent space, and a decoder; given input data x, the encoder output parameters are μ and log(σ). 2 Gaussian distribution Where z is the latent representation; Next, VAE sample z from The parameterization is redefined in the middle, where z = μ + σ ∈; the decoder is denoted as p. θ (x|z), reconstructing the compressed input data x; in this process, the reconstruction error is minimized using the evidence lower bound ELBO and the distribution of the latent space is regularized, defined as The first term is used for reconstruction, and KL(·) represents the divergence KL used for regularization; these two terms are used to achieve a balance between fitting the posterior and maintaining the latent space distribution; considering the substantial impact of significant noise on the latent space representation in the VAE, and this impact makes μ and log(σ) equal. 2 The estimated value of ) may become unstable, thus presenting an elusive latent variable z; in this case, the regularization term in formula (9) loses its effectiveness and cannot enforce the prior distribution, so the decoder abandons the use of the latent space, resulting in a generalized output x, characterized by posterior collapse; when facing weak signals, the approximate posterior distribution q φ (z|x) tends to approximate the prior distribution q φ (z), using This indicates that the KL divergence is invalid. In order to prevent the VAE from being trapped in the local optimum space, the Discrete Classification Variational Autoencoder (DC-VAE) is used. Learnable discrete vectors are used to form the latent category space. When predicting content popularity, the goal of DC-VAE is to find the vector with the closest encoding distance to the encoder network output in the latent category space, and then reconstruct the mapped vector through the decoder network. DC-VAE learns the latent distribution in the user request matrix x with the aim of obtaining future user requests in the reconstruction matrix output by the decoder; the user request matrix x contains historical information about the content of user requests at the edge nodes, defined as... In the formula, 1 ≤ n ≤ N, and n represents the number of users connected to a certain edge node; This represents the content request record for user n, where l ≤ i ≤ I, and i is the content library index; This indicates that the content request was successful; This indicates either a content request failure or a lack of interest in the content, but these two situations are difficult to distinguish, leading to inaccurate predictions. To address this issue, matrix X was supplemented and calibrated; specifically, in DC-VAE, the latent space is defined as... Where K is the size of the space and D is the dimension of the vector; therefore, there are K category vectors. DC-VAE inputs x through encoder network n Output t v (x n Next, the discrete latent variable t and the posterior probability distribution q(t|x) are calculated using the nearest neighbor algorithm. n ) is a one-hot encoding, defined as The input to the decoder is defined as t q (x n ) = v k Where k is the index of the decoder input, defined as k = argminj||t v (x n )-v j ||2; To address the gradient collapse problem caused by the introduction of implicit spaces, the gradient V is transformed during backpropagation. z L is copied from the decoder network to the encoder network; during training of the DC-VAE, the loss function is defined as... Where logp(xn|t) q (x n The reconstruction loss is used to optimize the encoder and decoder networks; since the backpropagation gradient is directly copied to the encoder network, the loss log p(t|t) is not considered. q (xn)); exist In the middle, the L2 error is used to measure v k Drive to t v (x n This aims to optimize the implicit category space; This is to prevent the encoder output from exceeding the range of the latent class space, where λ depends on the reconstruction loss, and sg is the stopping gradient operator, which is a constant during forward propagation and has a partial derivative of 0. Next, the log-likelihood function is defined as follows: logp(x n )≈logp(x n |t q (x n ))p(t q (x n )) (13) According to Jensen's inequality, equation (13) can be rewritten as follows: logp(x n )≥logpp(x n |t q (x n ))p(t q (x n )) (14)。 8. The multi-edge collaborative caching method based on robust federated deep learning according to claim 1, characterized in that: The residual-based detection specifically involves designing a parameter sorting matrix to prevent adversarial model updates from affecting the model's robustness. To detect adversarial updates, the model parameters from all edge nodes are first combined into a matrix. Defined as Among them, the global sharing model uses w (*) Parameterization, d m This represents local training data; Next, sort the elements in each column of R in descending order, preserving their sort order, and convert them to... Will The mean and variance var are defined as follows: Based on the mean and var, normal and adversarial model updates can be clustered into two groups using K-means. Adversarial model updates can be easily identified using the proposed residual-based detection method. Edge nodes providing normal model updates can be filtered using the proposed residual-based detection method, denoted as E. b ={e b，1 e b，2 , ..., e b，M′ } 9. The multi-edge collaborative caching method based on robust federated deep learning according to claim 8, characterized in that: The similarity-based federated aggregation uses canonical correlation analysis (CCA) to measure the similarity between the model update of each edge node and the average, thereby determining the weight of different model updates when performing federated aggregation. Described as Among them κ m This indicates the similarity score; The RFDL algorithm combines residual-based detection with similarity-based federated aggregation, and includes the following steps; Cloud data center update; First, set the total number r of FL communication wheels. max And initialize the global content popularity prediction model w (r) For each FL communication round, edge nodes update their local models in parallel; next, residual-based detection identifies adversarial model updates and obtains the edge node E that provides the normal model update. b Finally, a globally shared model is generated through similarity-based federated aggregation and distributed to edge nodes; Update at each edge node; first, initialize the training epoch c. max Batch size B and learning rate η; With the global sharing model w (r) The input is used by each edge node to begin its own local training; for each training epoch, DC-VAE uses the Adam optimizer in batches to train and update the local model; after local training, each edge node uploads its latest local model to the cloud data center; D. Active cache replacement with RFDL Based on the proposed RFDL, an active cache replacement strategy with multi-edge cooperation is designed; The key steps are as follows: For each edge node, initialize the cache space by partitioning the multi-dimensional cache space. temp And the user interval set H, when cache temp When the value is ≥0, RFDL is invoked to predict and rank content popularity, and content that the user is interested in will be placed in a temporary cache library C. temp In order to avoid cache redundancy caused by overlapping user-interested content in different intervals, C will be used. temp Replace with C s From C temp Select the current user range h s The most popular cache h Content; Next, delete the cache library C on each edge node. m The process involves identifying duplicate entries in the cache and updating the available cache space; this process is repeated until the cache space is fully utilized.

Images