Multi-edge collaborative caching method based on robust federal deep learning

Abstract

The invention provides a multi-edge collaborative caching method based on robust federal deep learning, which comprises the following steps: firstly, RoCoCache designs a novel division mechanism for a multi-dimensional caching space, and realizes accurate content recommendation in a user classification interval; and then, the RoCoCache develops a new discrete classification variational auto-encoder (DC-VAE), and the popularity degree of the content is accurately predicted by overcoming posterior collapse. And finally, the RoCoCache creates an original training mode and an active cache replacement strategy based on robust federal deep learning. According to the method, the antagonism model update detection based on the residual error and the federal aggregation based on the similarity are combined, so that the model damage caused by antagonism update is avoided, the active cache replacement can adapt to the optimized cache resources, and the cache performance is improved.

Description

Technical Field

[0001] The present 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 Art

[0002] With the significant development of 5G technology, more and more complex applications are being deployed in fields such as industrial manufacturing, digital economy, Internet of Vehicles, and smart cities. Taking cloud computing as an example, when computing, the tasks and data generated by applications are uploaded to the remote cloud for processing, causing serious 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 at the edge of the network closer to the terminal device, 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 is able to cache user-centric content within edge nodes, thereby reducing access latency and data duplication, thereby enhancing user experience and reducing system costs. However, cache performance is usually limited by the size and overhead of the edge cache space. Therefore, how to effectively utilize the cache space and improve cache performance has received widespread attention from academia and industry. In general, cache performance is constrained by multiple factors including cache size, content relevance, cache partitioning, and cache placement strategy. Previous studies have considered active content caching in a limited cache space based on user preferences. However, the cache space can be further partitioned based on user characteristics, activities, and request patterns. This approach optimizes the configuration of edge cache resources and is able to provide accurate content recommendations for similar users across multiple edge nodes. Exploring the potential relationship between user and content characteristics in a multi-dimensional space can improve the hit rate of user access 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 the edge node to which a user is connected does not match its request, the user can determine the content of its request from other edge nodes that perform collaborative caching. However, existing research has widely adopted deep reinforcement learning (DRL) and game theory methods. With the rapid increase in the number of terminal devices and the increasing sensitivity to user privacy, these methods may no longer be the best way to address security issues and challenges of suboptimal cache resource configuration. As a distributed training framework, federated learning (FL) is considered a promising solution to the above problems. Following its core principle, edge nodes collaborate to train a global model by uploading model parameters without leaking original data. FL maximizes the needs of heterogeneous users across multiple edge nodes through a well-trained global model while ensuring user privacy and scalability of large-scale network models. However, in complex MEC environments, unintentional model corruption or malicious adversarial model interference may lead to the inability to perform model training and degraded model update quality. For example, unintentional model corruption may occur due to noisy training labels, insufficient data samples, and low quality of uploaded models. In addition, malicious edge nodes may deliberately launch adversarial attacks to tamper with the model, such as Byzantine and backdoor attacks. In addition, the detection of adversarial attacks is computationally intensive because it requires a large amount of data collection and analysis to identify improper behavior. Studies have shown that running previous intrusion detection systems (IDS) on a 10MB dataset consumes more than 400 seconds of CPU time, 200 joules of energy, and 100MB of RAM on mobile devices, seriously affecting their normal operation. These problems not only exacerbate security challenges, but also degrade the performance of the global model, affect cache hit rates, and increase system overhead. Therefore, applying Robust Federated Deep Learning (RFDL) to edge caching is crucial to enhancing system security and optimizing cache performance.

[0004] Specifically, the main challenges in applying FL to deal with the multi-edge collaborative caching problem are summarized below:

[0005] 1) Discrete user feature distribution and diversity of requested content: Different users may have different content preferences due to their discrete characteristics. Therefore, it is challenging to find the potential relationship between discrete user feature distribution and diverse content requests.

[0006] 2) Model Scalability: As the number of terminal devices increases, there will be more data presenting discrete distribution, resulting in higher computational and communication overheads for caching scenarios with a single edge node. However, the classic centralized training framework suffers from limited model scalability.

[0007] 3) Model robustness: Edge nodes may inadvertently upload low-quality models or be subject to adversarial attacks from malicious nodes, resulting in severe degradation of robustness during model updates.

[0008] In order to address the above challenges, the present invention proposes a novel multi-edge collaborative caching framework based on robust federated deep learning (Collaborative Caching framework based on Robust federated deeplearning, RoCoCache) method. First, RoCoCache designs a novel partitioning mechanism for the multidimensional cache space, which achieves accurate content recommendation within the user classification interval. Next, RoCoCache develops a new discrete classification variational autoencoder (DC-VAE) to accurately predict the popularity of content by overcoming posterior collapse. Finally, RoCoCache creates an original training model and active cache replacement strategy based on robust federated deep learning. It is worth noting that the combination of residual-based adversarial model update detection and similarity-based federated aggregation avoids model destruction caused by adversarial updates, enables active cache replacement to adapt to optimized cache resources, and thus improves cache performance. Summary of the invention

[0009] The present 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 solutions.

[0011] A multi-edge collaborative caching method based on robust federated deep learning is used for a multi-edge collaborative caching system, wherein the multi-edge collaborative caching system consists of M edge nodes, with a set E = {e1, e2, ..., e m , ..., e M}, each node contains a MEC server, a base station, and N users, and the users are represented by the set U = {u1, u2, ..., u n , ..., u N} represents; the cache space of the edge node is defined as the set C = {C1, C2, ..., C m , ..., C M}; Each user is connected to an edge node and communicates through the wireless link provided by the relevant base station; In addition, the communication between edge nodes and between edge nodes and cloud data centers is carried out through backhaul links; The cache space state of each edge node is periodically broadcast to other edge nodes in 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; moreover, users are discretely distributed in the service areas of each edge node;

[0012] When user u n Send content f to the edge node it is connected to i When a request is received, a multi-edge collaborative caching method RoCoCache (Collaborative Caching framework based on Robustfederated deep learning, RoCoCache) for a multi-edge collaborative caching scenario is used, including the following steps:

[0013] Step S1: The current edge node checks whether f is cached i ; if f i is cached, the edge node sends it directly to u n ; Otherwise, go to step S2;

[0014] Step S2: The current edge node searches for cache f i If there is a cooperative edge node, the cooperative edge node will transmit f i Forward to the current edge node, and then f i Send to u n ; Otherwise, go to step S3;

[0015] Step S3: If no collaborative edge node caches f i , the content library in the cloud data center will provide f i , and forward it to the current edge node through 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 time-space dependencies. Improving cache hit rate relies on accurate content popularity prediction and caching the content that users are interested in into the cache space of edge nodes. Specifically, f i At the edge node e m The popularity on is defined as

[0017]

[0018] where req i，m for e m Received f i Number of requests, req m for e m The total number of requests received;

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

[0020]

[0021] Where r represents the FL communication wheel, w (r) is the parameter of the global prediction model, req is the total number of requests received by all edge nodes, is the model parameter of local FL, and the mean square error MSE is defined as

[0022]

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

[0024] And, define the cache hit ratio as

[0025]

[0026] where θ m (f i ) means e m Whether to cache the content requested by the user, defined as

[0027]

[0028] Cache performance is affected by multiple factors such as cache resource configuration, content popularity, model robustness, and cache replacement strategy. By comprehensively considering these factors, the multi-edge collaborative cache method RoCoCache can effectively improve the cache performance in the multi-edge collaborative cache system. The use of the multi-edge collaborative cache method RoCoCache method includes:

[0029] A. Through the multi-dimensional edge cache space partitioning mechanism, we can perceive and optimize cache resources to ensure that users receive accurate content recommendations within their classification intervals. This mechanism includes multi-dimensional user partitioning and cache space partitioning, taking into account differences in user characteristics, activities, and memory access intervals;

[0030] B. 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 from discrete vectors and then uses the nearest neighbor to find discrete latent vectors to assist the decoder in generating the user request matrix;

[0031] C. The RoCoCache method aggregates local models to generate a global shared model by training with user request data stored on each edge node; 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;

[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 through multi-dimensional cache space partitioning, and determines the appropriate cache space size for interval users;

[0033] Secondly, RoCoCache uses DC-VAE to learn the latent space containing discrete vectors to form a DC-VAE-based content popularity prediction. 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 adopts 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 a similarity-based FL aggregation method to avoid the damage caused by adversarial updates to the global 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, where the content that users are interested in is cached in 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 divergence of memory access intervals.

[0037] The multi-dimensional user segmentation is specifically as follows: assuming that user characteristics reflect their preferences for cached content to a certain extent; Figure 1 As shown in the figure, in order to accurately predict user preferences, RoCoCache uses a user segmentation method based on multidimensional features. The multidimensional features include gender, age, and occupation. These features are continuously encoded as coordinate axes, represented by the set Γ = {γ1, γ2, ..., γ t , ..., γ T}; it makes it possible to construct a feature hypercube, in which users are grouped to form user intervals; as the partitioning proceeds, the user interval is represented as a set H = {h1, h2, ..., h s , ..., h S}Where S is the number of user intervals, Grade(h s ) is the user classification level in the user interval; 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 evenly divided into 2 along each dimension. T intervals, where the length of each partition dimension will be halved, that is, ζ(Grade(h s )) determines the number of users in the user interval; when ζ(Grade(h s )) is larger, it will lead to a higher user density in each interval, thus affecting the ability to accurately capture the preferences of different users; on the contrary, a smaller ζ(Grade(h s )) values ​​correspond to reduced user density and may lead to inaccurate cache predictions;

[0038] In order to achieve adaptive segmentation of user intervals and capture the potential relationship between interval users and their preferred content, we set Where α is a hyperparameter. The division of each user interval may continue and enter the following stages according to performance requirements. The stage expression includes Grade = 1, 2, ...

[0039] The cache space division improves cache performance by allocating appropriate cache space to user intervals by considering multiple factors including the number of users, user activities, and the divergence of memory access intervals. Specifically,

[0040] In the divided user interval h s In the example, the number of users is recorded as |h s |. The divergence of user activity and memory access interval is defined as

[0041]

[0042] Where req s is the number of user requests, h s Memory access range.

[0043] Taking the above factors into consideration, the s The cache size is defined as

[0044]

[0045] In the formula, is the cumulative product function, cache m To connect h s The cache space size on the edge node.

[0046] The DC-VAE-based content popularity prediction is as follows: assume that the popularity of content in real scenarios usually follows a power-law distribution and always follows Zipf's law; use a variational autoencoder (VAE) in content popularity prediction to efficiently reconstruct compressed input data using continuous variables in the hidden layer so that the data is clustered in the latent space; specifically, the VAE consists of three main parts: an encoder, a latent space, and a decoder; in the case of input data x, the encoder output parameters are μ and log(σ 2 ) where z is the latent representation; next, VAE samples z from Reparameterize in, where z = μ + σ · ∈; the decoder is denoted by p θ (x|z), reconstruct the compressed input data x; in this process, the evidence lower bound ELBO is used to minimize the reconstruction error and regularize the distribution of the latent space, 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; if we consider the substantial impact of significant noise on the latent space representation in VAE, and this impact makes μ and log(σ 2 ) may become unstable, presenting an elusive latent variable z; in this case, the regularization term in formula (9) loses its effectiveness; it cannot enforce compliance with the prior distribution; the decoder abandons the use of the latent space, resulting in the output of a universal x, and the characteristic performance is the collapse of the posterior; when faced with weak signals, the approximate posterior distribution q φ (z|x) tends to be close to the prior distribution q φ (z), use In order to prevent VAE from being trapped in the local optimal solution space, a discrete classification variational autoencoder DC-VAE is used to form an implicit category space using a learnable discrete vector. When predicting the popularity of content, the goal of DC-VAE is to find the vector closest to the encoder network output encoding in the implicit category space, and then reconstruct the mapped vector through the decoder network.

[0049] Figure 2 shows the content popularity prediction based on DC-VAE; specifically, DC-VAE learns the implicit distribution in the user request matrix X, with the goal of obtaining future user requests in the reconstruction matrix output by the decoder; the user request matrix x contains the historical information of user requested content on edge nodes, defined as

[0050]

[0051] Where l≤n≤N, n represents the number of users connected to a certain edge node; represents the content request record of user n, where l≤i≤I, i is the content library index; Indicates that the content request was successful; Indicates that either the content request fails or the content is not interesting, but these two situations are difficult to distinguish, resulting in inaccurate predictions. To solve this problem, the matrix X is 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, DC-VAE inputs x through the encoder network n , output t v (x n );

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

[0053] The input of 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] In order to solve the gradient collapse problem caused by the introduction of implicit space, the gradient V is converted to z L is copied from the decoder network to the encoder network; when training DC-VAE, the loss function is defined as

[0055]

[0056] where log p(x n |t q (xn )) is the reconstruction loss, which aims to optimize the encoder and decoder networks; since the back-propagated gradient is directly copied to the encoder network, the loss log p(t|t q (x n ));

[0057] exist In the example, we use the L2 error to convert v k Drive to t v (x n ), which aims to optimize the latent category space; It is to prevent the encoder output from exceeding the range of the implicit category space, where λ depends on the reconstruction loss, sg is the stop gradient operator, which is a constant and the partial derivative is 0 during the forward propagation process;

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

[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, formula (13) can be rewritten as

[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 through parameter sorting; similarity-based federated aggregation is to avoid the damage of adversarial updates to the global shared model, and generate a robust and accurate content popularity prediction model in a complex multi-edge environment.

[0063] The residual-based detection is specifically as follows: To prevent adversarial model updates from affecting the robustness of the model, a parameter sorting matrix is ​​designed to detect adversarial updates; e.g. 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 shared model is w (*) Parameterization, d m Represents local training data;

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

[0067]

[0068] According to the mean and var, the normal and adversarial model updates can be divided into two clusters by K-means, where the adversarial model updates can be easily identified by the proposed residual-based detection. By using the proposed residual-based detection method, the edge nodes that provide 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 uses canonical correlation analysis (CCA) to measure the similarity between the model update of each edge node and the average, so as to determine the weights of different model updates when performing federated aggregation; described as

[0070]

[0071] where κ m represents the similarity score;

[0072] Combining residual-based detection with similarity-based federated aggregation, the RFDL algorithm is formed, which includes the following steps;

[0073] Update of cloud data center; First, set the total number of FL communication rounds r max And initialize the global content popularity prediction model w (r) ; For each FL communication round, the edge nodes update their local models in parallel; Next, residual-based detection is to identify adversarial model updates and obtain edge nodes E that provide normal model updates b ;Finally, a global shared model is generated through similarity-based federated aggregation and distributed to edge nodes;

[0074] Update at each edge node; first, initialize the training round c max , batch B and learning rate η; with a global shared model w (r)Each edge node starts its own local training with the input of . For each training round, 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 and 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 by multi-dimensional cache space partitioning. temp and user interval set H, when cache temp ≥0, RFDL is called to predict and sort the popularity of content, and the content that the user is interested in will be placed in the temporary cache library C temp In order to avoid cache redundancy caused by overlapping content of interest to users in different intervals, C temp Replace with C s , from C temp Select the current user interval h s The most popular cache h Next, delete the cache library C on each edge node. m The duplicates in the cache are found and the available cache space is updated; the above steps are repeated until the cache space is fully occupied.

[0076] This paper proposes a novel multi-edge collaborative caching frameworkbased on Robust federated deep learning (RoCoCache) method. First, RoCoCache designs a novel partitioning mechanism for the multidimensional cache space, which achieves accurate content recommendation within the user classification interval. Next, RoCoCache develops a new discrete classification variational autoencoder (DC-VAE) to accurately predict the popularity of content by overcoming posterior collapse. Finally, RoCoCache creates an original training model and active cache replacement strategy based on robust federated deep learning. It is worth noting that the combination of residual-based adversarial model update detection and similarity-based federated aggregation avoids model destruction caused by adversarial updates, enables active cache replacement to adapt to optimized cache resources, and thus improves cache performance.

[0077] The present invention designs RoCoCache, a resilient collaborative caching framework that uniquely integrates RFDL with active caching strategies. First, a multidimensional cache space partitioning mechanism is designed to optimize the edge cache space and provide accurate content recommendations within user interval classification. Next, a content popularity prediction algorithm based on DC-VAE is developed to solve the posterior collapse problem and improve the prediction accuracy. Finally, an RFDL-based training mode and active cache replacement strategy are established to improve adaptability and robustness in complex network environments. Using real-world test platforms and datasets, it is demonstrated that RoCoCache achieves a higher cache hit rate than the most advanced methods and is close to the optimal hit rate. Through ablation experiments, it is verified that the design of multidimensional cache space partitioning and collaborative caching can improve cache performance. In addition, RoCoCache shows excellent training and caching efficiency in various scenarios. In addition, RoCoCache is able to identify adversarial model updates in complex network environments, verifying its good robustness. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0079] Attached Figure 1 It is a schematic diagram of multi-dimensional cache space division in the present invention;

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

[0081] Attached Figure 3 is a schematic diagram of the RFDL model in the present invention;

[0082] Attached Figure 4 is a schematic diagram comparing the embodiment with the State-of-the-Arts method;

[0083] Attached Figure 5 is a comparative schematic diagram of ablation experiments in the embodiment;

[0084] Attached Figure 6 is a comparative schematic diagram of convergence analysis in an embodiment;

[0085] Attached Figure 7 is a comparative schematic diagram of training efficiency in an embodiment;

[0086] Attached Figure 8 is a comparative schematic diagram of model detection in an embodiment;

[0087] Attached Fig. 9 is a schematic diagram of comparison of robust performance in an embodiment; DETAILED DESCRIPTION

[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, including a workstation and 20 Jetson TX2s. The workstation serves as a cloud data center, equipped with two NVIDIA GeForce GTX 3090 GPUs, an Intel (R) Xeon (R) CPU Silver 4208 @ 2.10GHz and 32GB RAM. Jetson TX2s act as edge nodes, each equipped with an NVIDIA Pascal GPU with 256 CUDA-capable cores, and a CPU processor consisting of a 2-core Denver2 and a 4-core ARM Cortex A57. The workstation and Jetson TX2s are deployed in the same network. Based on the FLASK web framework, a backend service is built to serve the communication between the workstation and Jetson TX2s. The above testbed devices use the Ubuntu 18.04 operating system, CUDA v10.0 and cuDNN v7.5.0.

[0091] Datasets. Three real-world datasets, MovieLens 100K and 1M, and iQiyi, are used. These three datasets contain approximately 100,000, 1,000,000, and 10,000,000 request records, respectively. Among them, the datasets of MovieLens 100K and 1M are collected by groupLens Research. MovieLens 100K contains approximately 10,000 ratings from 943 anonymous users on 1,682 movie items, and MovieLens 1M contains approximately 1,000,000 ratings from 6,040 anonymous users on 3,883 movie items. In addition, the iQiyi dataset released by the iQiyi website contains approximately 10 million requests from 100,000 users for 1,900,000 video items. These datasets provide user sequence numbers, item indexes, timestamp labels, and user context information. Specifically, the user's gender, age, and occupation are selected as user features, and the ratings are regarded as user requests. The dataset is divided into training set (70%), validation set (10%) and test set (20%).

[0092] Parameter setting. Based on the above real-world testbed and dataset, a multi-edge collaborative caching scenario consisting of a cloud data center, 5 to 20 edge nodes, and 9.43 million to 100,000 users is simulated. 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 in the service area of ​​each edge node. RoCoCache is implemented based on Python 3.8 and Tensorflow 2.4.0. Specifically, the hyperparameter α in the multidimensional cache space partition is 512, the size of the DC-VAE hidden category space K is 128, and the dimension of the category vector Dv is c is 16, the number of FL communication rounds is r max is 30, the batch size in DC-VAE is 32, and the number of training epochs is c max is 300, and the learning rate η is 0.001. According to experimental tests, the content request delays for accessing local edge nodes, collaborative edge nodes, and cloud data centers are around [2, 4] ms, [15, 20] ms, and [18, 23] ms, respectively.

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

[0094] (1) Oracle: Knows all the prior information about future user requests, so it can achieve the optimal cache hit rate under limited cache space.

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

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

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

[0098] (5) Edge Cooperative Cache (ECC): Integrates a neural collaborative 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 using the encoder, and then the predicted content popularity distribution is obtained from the output matrix.

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

[0101] Attack Model. The robustness of RoCoCache is evaluated by using the following two attack models.

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

[0103] Gaussian Noise Attack (GNA): Generate 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 are conducted on various datasets with different edge cache space sizes. Figure 4 As shown in the figure, as the size of the edge cache space increases, the cache hit rates of all methods show an increasing trend. Since Oracle knows the prior information of future user requests, it achieves the theoretically optimal result. Random Cache shows the worst performance because its cache strategy is blind. AE and VAE show good cache hit rates because they compress high-dimensional user requests into low-dimensional representations and learn the potential relationship between user features and request content. However, due to the difference between the large data volume and the single distribution assumption, AE suffers from the posterior collapse problem on the iQiyi dataset. Therefore, AE shows an extremely low cache hit rate. By using clustering in a continuous latent space, VAE has a better ability to reconstruct the input distribution than AE, so VAE can obtain more accurate content popularity prediction and higher cache hit rate. In addition, as the edge cache space increases, ECC shows a decreasing trend in the growth of cache hit rate, and its performance is always inferior to RoCoCache. This is because the collaborative filter adopted by ECC cannot accurately capture content popularity, resulting in inefficient utilization of edge cache space. LECS adopts a content popularity prediction model based on a temporal convolutional network and achieves a growth rate comparable to other popularity-driven methods. However, when content popularity exhibits similar patterns but changes at different amplitudes, it is difficult for such a temporal neural network to establish clear boundaries for effective caching, which ultimately reduces the cache hit rate. LRU can respond well to bursty and sparse content requests, but it is difficult to handle the changing trend of content popularity. Therefore, the cache performance of LRU is inferior to that of AE, VAE, ECC, and LECS. Compared with other methods, RoCoCache shows a higher cache hit rate, which is close to the best. This is because RoCoCache achieves perceptual optimization of edge cache space through multi-dimensional cache space partitioning, while solving the posterior collapse problem in VAE, making content popularity prediction more accurate.

[0106] Ablation experiments. Ablation experiments are conducted on the MovieLens 1M dataset to evaluate the effectiveness of the multi-dimensional cache space partitioning and collaborative caching designed in RoCoCache. Figure 5 As shown in Figure 3, the cache hit rates of all methods increase 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 collaborative caching, its performance is comparable to RoCoCache, which only uses multi-dimensional cache space partitioning. When RoCoCache adopts collaborative caching, it can benefit from the joint optimization of multi-edge cache resources, thereby achieving a higher cache hit rate. It is worth noting that RoCoCache obtains the best cache hit rate in all ablation experiments, which indicates that multi-dimensional cache space partitioning can effectively assemble the common user preferences of each edge node. Therefore, RoCoCache maintains good caching performance in both collaborative and non-collaborative caching scenarios. The hit rate of RoCoCache under attack is closely consistent with that of RoCoCache, which will be thoroughly analyzed in the robustness analysis section.

[0107] Convergence analysis. Figure 6 The convergence of RoCoCache with the increase of FL communication rounds on the MovieLens 1M dataset is analyzed. Different sizes of edge cache spaces are considered when conducting the experiment to ensure the comprehensiveness of the test. In the initial stage, content is randomly selected and stored in the cache space of the edge node, resulting in a low cache hit rate. After one round of FL communication, RoCoCache generates a preliminary prediction model for content popularity, thereby quickly improving the cache hit rate. At this time, in scenarios with different edge cache space sizes, RoCoCache can achieve more than 80% of the optimal cache performance, where the size of the edge cache space determines the growth range of the cache hit rate. It is worth noting that in different scenarios, RoCoCache has a tendency to converge after 6 rounds of FL communication, which illustrates the excellent convergence of RoCoCache.

[0108] Training efficiency. The training efficiency of RoCoCache on the MovieLens 1M dataset was evaluated in scenarios with different numbers of edge nodes. Figure 7As shown in the figure, the time of each round of FL training decreases as the number of edge nodes increases. When the number of user content requests is constant, adding edge nodes for collaborative caching can effectively improve the training efficiency of RoCoCache. In addition, as the number of edge nodes increases, RoCoCache can better capture different user preferences and improve the cache hit rate through multi-edge collaboration. The experimental results show that RoCoCache can adapt to various multi-edge scenarios and has excellent training efficiency and cache hit rate.

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

[0110]

[0111] Cache efficiency. The cache efficiency of different methods in 10 edge node scenarios was tested in terms of the latency of content requests. Non-collaborative means RoCoCache without collaborative caching, and distributed means only one copy of the content is cached on each edge node based on the popularity of the content. As shown in Table 5-1, as the size of the edge cache space increases, the latency of content requests decreases. RoCoCache can handle content requests in different ways and accurately predict the popularity of content, so it has the best cache efficiency. Non-collaborative does not use collaborative caching, so requests for missing content need to be forwarded to the remote cloud, resulting in lower cache efficiency. Moreover, due to the low cache hit rate, distributed needs to continuously send content requests to other edge nodes and the remote cloud. Therefore, compared with RoCoCache, the other two methods will cause excessive latency.

[0112] Robustness Analysis. The robustness of RoCoCache is evaluated from two aspects. First, the ability of RoCoCache to detect adversarial model updates is tested. Figure 8 The performance of residual-based detection models is shown when exposed to 30% and 40% of model updates being adversarial. The separation between adversarial and normal model updates shows the different means and variances of the two update types in terms of gradient parameters. For detecting SFA, the increase in the proportion of adversarial model updates increases the difficulty of using residual-based detection. In this case, RoCoCache can still distinguish adversarial model updates. For detecting GNA, a larger proportion of Gaussian noise severely affects the variance of the parameter ranking matrix, and the separation is more obvious.

[0113] Secondly, Fig. 9The robustness performance of RoCoCache under different attacks and defenses is described when the adversarial model update ratio is 30%. In the absence of attacks, RoCoCache can achieve an ideal cache hit rate. Under the attacks of SFA and GNA, RoCoCache can still converge after about 20 rounds of FL communication, which is close to the ideal result. In the absence of defense, SFA and GNA will seriously affect the cache performance. In this case, it is obvious that the adjustment of the cache strategy is unlikely to significantly improve the cache hit rate. SFA leads to the sign reversal problem, severe joint aggregation and rendering the global shared model invalid. At the same time, GNA seriously affects the cache performance in the absence of defense. This is because Gaussian noise changes the weighted mean and geometric median of the global shared model, increasing the difficulty of model training. The results verify the strong robustness of RoCoCache, provide accurate identification of adversarial model updates in complex network environments, and ensure the rapid convergence of the model.

Claims

1. A multi-edge collaborative caching method based on robust federated deep learning, used in a multi-edge collaborative caching system, characterized by: The multi-edge collaborative cache system is composed of M edge nodes, and the set E = {e1, e2, ..., e m , ..., e M } indicates that each node contains a MEC server, a base station, and N users, and the users are represented by the set U = {u1, u2, ...., u n , ..., u N } represents; the cache space of the edge node is defined as the set C = {C1, C2, ..., C m , ..., C M }; Each user is connected to an edge node and communicates through the wireless link provided by the relevant base station; In addition, the communication between edge nodes and between edge nodes and cloud data centers is carried out through backhaul links; The cache space state of each edge node is periodically broadcast to other edge nodes in 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; moreover, users are discretely distributed in the service areas of each edge node; When user u n Send content f to the edge node it is connected to i When a request is made, the multi-edge collaborative caching method RoCoCache is used for multi-edge collaborative caching scenarios. The steps include: Step S1: The current edge node checks whether f is cached i ; if f i is cached, the edge node sends it directly to u n ; Otherwise, go to step S2; Step S2: The current edge node searches for a cooperating edge node with a cache. If so, the cooperating edge node will send the cached data to the cooperating edge node through the backhaul link. i Forward to the current edge node, and then f i Send to u n ; Otherwise, go to step S3; Step S3: If no collaborative edge node caches f i , the content library in the cloud data center will provide f i , and forward it to the current edge node through the backhaul link, and then f i Send to u n .

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 temporal and spatial dependencies; Improving cache hit rate depends on accurate content popularity prediction and caching the content that users are interested in into the cache space of edge nodes; Specifically, f i At the edge node e m The popularity on is defined as where req i，m for e m Received f i Number of requests, req m for e m The total number of requests received; The method RoCoCache accurately predicts the popularity of content; to evaluate the prediction accuracy, the global loss function is defined as Where r represents the FL communication wheel, w (r) is the parameter of the global prediction model, req is the total number of requests received by all edge nodes, is the model parameter of local FL, and the mean square error MSE is defined as in f i For m The predicted popularity value, P i，m (r) is the actual value; And, define the cache hit ratio as where θ m (f i ) means e m Whether to cache the content requested by the user, defined as Cache performance is affected by many factors such as 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 in the multi-edge collaborative caching system.

3. The multi-edge collaborative caching method based on robust federated deep learning according to claim 2, characterized in that: The multi-edge collaborative caching method RoCoCache method is used in the following ways: A. Through the multi-dimensional edge cache space partitioning mechanism, we can perceive and optimize cache resources to ensure that users receive accurate content recommendations within their classification intervals. This mechanism includes multi-dimensional user partitioning and cache space partitioning, taking into account differences in user characteristics, activities, and memory access intervals; B. 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 from discrete vectors and then uses the nearest neighbor to find discrete latent vectors to assist the decoder in generating the user request matrix; C. The RoCoCache method aggregates local models to generate a global shared model by training with user request data stored on each edge node; 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; 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 through multi-dimensional cache space partitioning, and determines the appropriate cache space size for interval users; Secondly, RoCoCache uses DC-VAE to learn the latent space containing discrete vectors to form a DC-VAE-based content popularity prediction. 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 adopts 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 a similarity-based FL aggregation method to avoid the damage caused by adversarial updates to the global 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.

4. The multi-edge collaborative caching method based on robust federated deep learning according to claim 3 is characterized in that: The multi-dimensional cache space partitioning includes two key components: multi-dimensional user partitioning and cache space partitioning; in multi-dimensional user partitioning, RoCoCache classifies and segments feature groups with different numbers of users, where the content that users are interested in is cached in different groups; In cache space partitioning, the established classification is utilized to perform perceptual optimization of cache space based on user characteristics, user activities, and the divergence 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 multi-dimensional user segmentation is specifically as follows: assume that user characteristics reflect their preferences for cached content to a certain extent; in order to accurately predict user preferences, RoCoCache uses a user segmentation method based on multi-dimensional features, where the multi-dimensional features include gender, age, and occupation. These features are continuously encoded as coordinate axes, using the set Γ = {γ1, γ2, ..., γ t , ..., γ T }; it makes it possible to construct a feature hypercube, in which users are grouped to form user intervals; as the partitioning proceeds, the user interval is represented as a set H = {h1, h2, ···, h s ,···,h S }, where S is the number of user intervals, Grade(h s ) is the user classification level in the user interval; 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 evenly divided into 2 along each dimension. T intervals, where the length of each partition dimension will be halved, that is, Determines the number of users in the user interval; when ζ(Grade(h s )) is larger, it will lead to a higher user density in each interval, thus affecting the ability to accurately capture the preferences of different users; on the contrary, a smaller ζ(Grade(h s )) values ​​correspond to reduced user density and may lead to inaccurate cache predictions; In order to achieve adaptive segmentation of user intervals and capture the potential relationship between interval users and their preferred content, we set Where α is a hyperparameter. The division of each user interval may continue and enter the following stages according to performance requirements. The stage expression 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 division improves cache performance by allocating appropriate cache space to user intervals by considering multiple factors including the number of users, user activities, and the divergence of memory access intervals. Specifically, In the divided user interval h s In the example, the number of users is recorded as |h s |. The divergence of user activity and memory access interval is defined as Where req s is the number of user requests, h s Memory access range. Taking the above factors into consideration, the s The cache size is defined as In the formula, is the cumulative product function, cache m To connect h s The cache space size on the edge node.

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 is specifically as follows: assuming that the popularity of content in real scenarios usually follows a power law distribution and always follows Zipf's law; using a variational autoencoder VAE in content popularity prediction to efficiently reconstruct the compressed input data using continuous variables in the hidden layer, so that the data is clustered in the latent space; specifically, VAE consists of three main parts: an encoder, a latent space, and a decoder; in the case of input data x, the encoder output parameters are μ and log(σ 2 ) where z is the potential representation; Next, the VAE samples z from Reparameterize in, where z = μ + σ · ∈; the decoder is denoted by p θ (x|z), reconstruct the compressed input data x; in this process, the evidence lower bound ELBO is used to minimize the reconstruction error and regularize the distribution of the latent space, 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. If we consider the substantial impact of significant noise on the latent space representation in VAE, and this impact makes μ and log(σ 2 ) may become unstable, presenting an elusive latent variable z; in this case, the regularization term in formula (9) loses its effectiveness; it cannot enforce compliance with the prior distribution; the decoder abandons the use of the latent space, resulting in the output of a universal x, and the characteristic performance is the collapse of the posterior; when faced with weak signals, the approximate posterior distribution q φ (z|x) tends to be close to the prior distribution q φ (z), use In order to prevent VAE from being trapped in the local optimal solution space, a discrete classification variational autoencoder DC-VAE is used to form an implicit category space using a learnable discrete vector. When predicting the popularity of content, the goal of DC-VAE is to find the vector closest to the encoder network output encoding in the implicit category space, and then reconstruct the mapped vector through the decoder network. DC-VAE learns the implicit distribution in the user request matrix X, with the goal of obtaining future user requests in the reconstruction matrix output by the decoder; the user request matrix X contains the historical information of the user request content on the edge nodes and is defined as Where 1≤n≤N, n represents the number of users connected to a certain edge node; represents the content request record of user n, where l≤i≤I, i is the content library index; Indicates that the content request was successful; Indicates that either the content request fails or the content is not interesting, but these two situations are difficult to distinguish, resulting in inaccurate predictions. To solve this problem, the matrix X is 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 the encoder network n , output t v (x n ); Next, the discrete latent variable t is calculated by the nearest neighbor algorithm, and the posterior probability distribution q(t|x n ) is a one-hot encoding, defined as The input of 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; In order to solve the gradient collapse problem caused by the introduction of implicit space, the gradient V is converted to z L is copied from the decoder network to the encoder network; when training DC-VAE, the loss function is defined as where log p(x n |t q (x n )) is the reconstruction loss, which aims to optimize the encoder and decoder networks; since the back-propagated gradient is directly copied to the encoder network, the loss log p(t|t q (x n )); exist In the example, we use the L2 error to convert v k Drive to t v (x n ), which aims to optimize the latent category space; It is to prevent the encoder output from exceeding the range of the implicit category space, where λ depends on the reconstruction loss, Sg is the stop gradient operator, which is a constant and the partial derivative is 0 during the forward propagation process; Next, the log-likelihood function is defined as log p(x n )≈log p(x n |t q (x n ))p(t q (x n )) (13) According to Jensen's inequality, formula (13) can be rewritten as log p(x n )≥log p(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 4, characterized in that: The robust federated deep learning RFDL includes residual-based detection and similarity-based federated aggregation; residual-based detection detects adversarial model updates through parameter sorting; similarity-based federated aggregation is to avoid the damage of adversarial updates to the global shared model, and generate a robust and accurate content popularity prediction model in a complex multi-edge environment.

9. The multi-edge collaborative caching method based on robust federated deep learning according to claim 8, characterized in that: The residual-based detection is specifically as follows: To prevent adversarial model updates from affecting the robustness of the model, a parameter sorting matrix is ​​designed To detect adversarial updates; first combine the model parameters from all edge nodes into a matrix Defined as Among them, the global shared model is w (*) Parameterization, d m Represents local training data; Next, sort the elements of each column in R into descending order, preserving their sorted position, and convert them to Will The mean and variance var are defined as According to the mean and var, the normal and adversarial model updates can be divided into two clusters by K-means, where the adversarial model updates can be easily identified by the proposed residual-based detection. By using the proposed residual-based detection method, the edge nodes that provide normal model updates can be filtered, denoted as E b ={e b，1 , e b，2 , ..., e b，M′ }.

10. 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, so as to determine the weights of different model updates when performing federated aggregation; Described as where κ m represents the similarity score; Combining residual-based detection with similarity-based federated aggregation, the RFDL algorithm is formed, which includes the following steps; Update of cloud data center; First, set the total number of FL communication rounds r 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 is to identify adversarial model updates and obtain edge nodes E that provide normal model updates b ;Finally, a global shared model is generated through similarity-based federated aggregation and distributed to edge nodes; Update at each edge node; first, initialize the training round c max , batch B and learning rate η; With a global shared model w (r) Each edge node starts its own local training with the input of . For each training round, 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 and RFDL Based on the proposed RFDL, a proactive cache replacement strategy with multi-edge cooperation is designed; The key steps are as follows: For each edge node, initialize the cache space cache by multi-dimensional cache space partitioning temp and user interval set H, when cache temp ≥0, RFDL is called to predict and sort the popularity of content, and the content that the user is interested in will be placed in the temporary cache library C temp In order to avoid cache redundancy caused by overlapping content of interest to users in different intervals, C temp Replace with C s , from C temp Select the current user interval h s The most popular cache h Next, delete the cache library C on each edge node. m The duplicates in the cache are found and the available cache space is updated; the above steps are repeated until the cache space is fully occupied.

Images