A federated learning assisted edge caching method based on AE-DDPM model

Abstract

The application relates to a federated learning assisted edge caching method based on an AE-DDPM model, and comprises the following steps: an edge computing network system model is established, including a base station, a remote cloud server and users; an AE-DDPM model based on federated learning is used for training; global prediction content popularity is obtained, and the most popular N contents are cached according to the cache capacity of the base station; the application extracts a user data potential feature vector through an AE model, reduces data dimension and sparseness, and then learns data distribution through a DDPM model to generate high-quality content popularity prediction; the application deploys a cache unit on an edge node, enables users to quickly obtain the pre-cached contents on the node, effectively improves the cache hit rate, reduces the time delay of the users in obtaining the contents, significantly improves communication efficiency, and simultaneously reduces the risk of user privacy leakage.

Description

Technical Field

[0001] This invention relates to the field of edge caching technology, and in particular to a federated learning-assisted edge caching method based on the AE-DDPM model. Background Technology

[0002] In recent years, with the rapid development of the mobile internet and the increasing popularity of smart devices, mobile data traffic has experienced explosive growth, putting enormous pressure on wireless networks. Against this backdrop, edge caching technology has emerged. Edge caching technology pre-stores content that users may be interested in by deploying caching units at the network edge, such as base stations and access points. When users request this content, it can be retrieved directly from nearby edge nodes without having to download it from distant data centers or the cloud, thereby reducing content distribution latency, alleviating network congestion, improving user experience, and enhancing overall system performance.

[0003] To effectively enable users to access content of interest from nearby base stations, base stations must predict popular content based on user preferences within their coverage area. Machine learning techniques can extract latent features from user data and predict preferred content. However, user personal data often contains a large amount of privacy-sensitive information, and users typically refuse to share their data directly with others. Traditional centralized machine learning methods require collecting user data on a central server for training, which not only risks infringing on user privacy but also faces problems such as data leakage and high transmission costs. Federated learning solves this problem. As an emerging distributed machine learning method, federated learning enables multiple participants to collaboratively train machine learning models while protecting user privacy. By sharing local model parameters and other information instead of raw data, federated learning avoids the risk of leaking user privacy information, providing a reliable way to train models in scenarios where data is widely distributed and privacy is sensitive, such as mobile devices and IoT devices.

[0004] While traditional federated learning methods can protect user privacy to some extent, they suffer from low communication efficiency. The participants need to communicate frequently and the data volume is enormous, which leads to increased network congestion and latency. Furthermore, traditional federated methods are limited in model performance when dealing with sparse data, making it difficult to accurately extract effective features and affecting the accuracy of content preference prediction. Summary of the Invention

[0005] In view of this, the present invention provides a federated learning-assisted edge caching method based on the AE-DDPM model, which can more accurately predict the content that users are interested in while protecting user privacy, and cache the predicted content on the base station, thereby effectively reducing the latency for users to obtain the content they are interested in.

[0006] To achieve the above objectives, this invention provides a federated learning-assisted edge caching method based on the AE-DDPM model, comprising the following steps:

[0007] S1. Establish an edge computing network system model, including base stations, remote cloud servers, and users;

[0008] S2. The AE-DDPM model based on federated learning is used for training;

[0009] S201, Base station generates global initial AE-DDPM model ω 0 The global AE-DDPM model ω 0 Distribute to users for training;

[0010] The S202 and AE models are trained locally for e iterations, extracting latent feature vectors from the user's training data. And used for e iterations of training the DDPM model;

[0011] The process of training the AE model includes:

[0012] In the k-th iteration, user i starts from the local training set B i Randomly select a subset

[0013] The subset The training data points z are input into the AE model, and the AE model generates reconstructed output data points. In the k-th iteration, the local AE model is updated as follows:

[0014]

[0015] in, This represents the AE model parameters for user i in the r-th round and k-th iteration. express gradient, Let η represent the local loss function of the AE model in the k-th iteration. a This represents the learning rate of the AE model;

[0016] The training data after e iterations of AE model training is input into the encoder network of the AE model to obtain the latent feature vector. And input it into the DDPM model for e iterations of training;

[0017] The process of training the DDPM model includes:

[0018] In the k-th iteration, user i starts from the local training set. Randomly select a subset

[0019] In the k-th iteration, the local DDPM model is updated as follows:

[0020]

[0021] in, This represents the DDPM model parameters for user i in the r-th round and k-th iteration. express gradient, Let η represent the local loss function of the k-th iteration of the DDPM model. d This represents the learning rate of the DDPM model;

[0022] After completing e iterations of DDPM model training, the r-th round of local training ends, yielding the r-th round AE-DDPM model ω. r :

[0023] ω r ={ω r,a ,ω r,d}

[0024] Where, ω r,a ω r,d Let represent the parameters of the AE model and the DDPM model in the r-th round, respectively;

[0025] S203, The user will update the r-th round of the AE-DDPM model ω r Uploaded from the local server to the base station;

[0026] S204. The base station calculates the weighted sum of the AE-DDPM models of all users within its coverage area to obtain a new global AE-DDPM model ω. r+1 ;

[0027] S205, the new global AE-DDPM model ω r+1 Used for the next round of training, when the number of training rounds reaches a preset threshold R. max Training ends, and the final AE-DDPM model is obtained;

[0028] S3. Obtain the global predicted content popularity and cache the N most popular contents based on the base station's cache capacity.

[0029] Preferably, the data points generated by the AE model reconstruction The expression is:

[0030]

[0031] in, Let D(·) represent the AE model parameters of user i in the r-th round and k-th iteration, and let D(·) and E(·) represent the decoder and encoder of the AE model, respectively.

[0032] The loss function expression for the data point Z is:

[0033]

[0034] The local loss function expression for the k-th iteration of the AE model is:

[0035]

[0036] in, Representing a subset The quantity and size value.

[0037] Preferably, the latent feature vector The expression is:

[0038]

[0039] in, Let B represent the AE model parameters for user i in the r-th iteration. i This represents the training set.

[0040] Preferably, the local loss function of the DDPM model The expression is:

[0041]

[0042] in, Representing data loss function, Representing a subset One of the data points, Let represent the DDPM model parameters for user i in the r-th round and k-th iteration.

[0043] Preferably, the new global AE-DDPM model ω r+1 The expression is:

[0044]

[0045] Where, d i d represents the local server data size of user i, d represents the total data size of all users within the base station coverage area, and η represents the learning rate of model aggregation.

[0046] Preferably, predicting content popularity includes the following steps:

[0047] The base station uses the global DDPM model ω dPerform a backdiffusion process and generate U spurious samples g. fake,u ;

[0048] U of the aforementioned fake samples g fake,u Input AE model decoder network to generate reconstructed fake samples The expression is:

[0049]

[0050] Where, ω a D represents the global AE model, D(·) represents the decoder of the AE model, and U represents the number of fake samples;

[0051] The reconstructed fake sample Instead of user data, content scores are calculated to predict content popularity in base stations.

[0052] Preferably, calculating the content score includes the following steps:

[0053] The reconstructed fake sample By adding the dimensions together, we can obtain a score for all content. The expression is:

[0054]

[0055] Where F represents the number of content types contained in the content library, and F represents the reconstructed fake sample. The higher the score, the more popular the content.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] This invention extracts latent feature vectors from user data using the AE model, reducing data dimensionality and sparsity, and providing high-quality input for the DDPM model. DDPM generates the data distribution required for content popularity prediction through progressive denoising, improving prediction accuracy and overcoming the limitations of traditional federated learning in handling sparse data. At the same time, this invention can more accurately predict content popularity, enabling base stations to cache content that better meets user needs, significantly improving cache hit rate, reducing latency for users to obtain content, enhancing user experience and overall system performance, and further reducing the risk of user privacy leakage. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of the scenario of the present invention;

[0059] Figure 2 A graph showing the change in cache hit rate as a function of cache capacity for different caching methods;

[0060] Figure 3A graph showing how latency for requesting content changes with cache capacity for different caching methods;

[0061] Figure 4 This is a graph showing the relationship between cache hit rate and request content latency in this invention.

[0062] Figure 5 This is a graph showing the change in cache hit rate of the present invention as a function of the number of users participating in training.

[0063] Figure 6 This is a graph showing the change in cache hit rate of different models in this invention with the number of training rounds. Detailed Implementation

[0064] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0065] Existing federated methods struggle to accurately extract effective features when dealing with sparse data, impacting the accuracy of content preference prediction. Autoencoders (AEs) are unsupervised learning neural network models, primarily composed of an encoder and a decoder: the encoder maps input data to a low-dimensional latent space, extracting feature vectors; the decoder reconstructs the original data based on these latent feature vectors. Autoencoders have wide applications in feature extraction, dimensionality reduction, and image generation, learning efficient representations of data and providing better input features for machine learning tasks. The Denoising Difficult Probabilistic Model (DDPM) is a generative model based on a latent diffusion process. Its core idea is to progressively add noise to the data, gradually transforming the data distribution into a noise distribution, and then training a neural network to learn the inverse denoising process to recover the original data from the noise. DDPM models have demonstrated strong performance in tasks such as image generation, speech synthesis, and data interpolation, generating high-quality samples.

[0066] To handle sparse user data, this embodiment provides a federated learning-assisted edge caching method based on the AE-DDPM model. The combination of the AE and DDPM models can predict content popularity while protecting user privacy, and includes the following steps:

[0067] S1, such as Figure 1 As shown, an edge computing network system model scenario diagram is established, including a base station, a remote cloud server, and I users;

[0068] The base station and the remote server are connected via a reliable backhaul link. All users are within the coverage area of ​​the base station, and each user i has a smart device. The base station has limited storage capacity and can store a maximum of N content items at the same time, while the remote cloud server caches all available content. When the content requested by the user is already cached at the base station, the base station will directly deliver the content to the user. Otherwise, the base station will request the content from the remote cloud server before delivering it to the user, which will result in a higher content request latency.

[0069] The communication model used in this embodiment is as follows: multiple users communicate with the base station in the sub-6GHz band using Orthogonal Frequency Division Multiplexing (OFDM). Due to the orthogonality of OFDM subcarriers and the frequency domain resource allocation mechanism, it is assumed that the system can avoid interference between users through optimized resource scheduling; the transmission rate R between the base station and user i is... s,i The expression is:

[0070]

[0071] Among them, B s,i P represents the available communication bandwidth between the base station and user i. s This indicates the base station's transmit power. Indicates the noise power at the receiver, g i H represents the small-scale fading factor, which follows an exponential distribution with a mean of 1. i This represents the large-scale fading factor, including the effects of shading and path loss; shading follows a log-normal distribution, and the path loss expression is:

[0072] P loss =20log 10 (dis(s,i))+20log 10 (f)+32.4

[0073] Where dis(s,i) represents the straight-line distance between the base station and user i, and f represents the carrier frequency in megahertz;

[0074] S2. The AE-DDPM model based on federated learning is used for training;

[0075] The theoretical basis of the diffusion model originates from the entropy increase-reverse process of non-equilibrium thermodynamic systems. The denoised diffusion probability model DDPM realizes the forward diffusion process and the reverse diffusion process through a parameterized Markov chain.

[0076] Forward diffusion process: by using a scheduling strategy A parameterized Markov chain is used, and Gaussian noise is gradually added to the original data, causing the data distribution to be gradually perturbed into random noise. The expression for the single-step diffusion process is:

[0077]

[0078] Where, x t This represents the data from the diffusion process at step t. Represents a Gaussian distribution, β t Parameters indicating the control of noise levels;

[0079] A linearly varying scheduling strategy is adopted, where From β t =10 -4 linearly increase to β T =0.02;

[0080] definition:

[0081]

[0082] Then we can get:

[0083]

[0084] Backdiffusion process: through training a neural network μ θ Predict the noise at each step ∈ θ To recover the original data distribution from noisy data; parameterized by conditional probability, the expression is:

[0085]

[0086] A simplified optimization objective function is obtained, expressed as:

[0087]

[0088] S201, Model Download;

[0089] The base station generates a global initial AE-DDPM model ω 0 The AE-DDPM model consists of AE and DDPM; ω r ω represents the global AE-DDPM parameters in the r-th training round. For subsequent rounds, the base station updates the global model at the end of the previous training round. r,a ω r,d Let represent the parameters of AE and DDPM in the r-th round, respectively, therefore ω r ={ω r,a ,ω r,d}; The global AE-DDPM model ω 0 Distribute to users for training;

[0090] S202, Local Training, including training the AE model, data processing, and training the DDPM model; the AE model undergoes e iterations of training locally, extracting latent feature vectors from the user's training data. The extracted data was then used for e-th iteration training of the DDPM model;

[0091] The process of training the AE model includes:

[0092] In the k-th iteration, user i starts from the local training set B i Randomly select a subset

[0093] subset The training data points z are input into the AE model, and the AE model generates reconstructed output data points. The expression is:

[0094]

[0095] in, Let D(·) represent the AE model parameters of user i in the r-th round and k-th iteration, and let D(·) and E(·) represent the decoder and encoder of the AE model, respectively.

[0096] The loss function expression for data point z is:

[0097]

[0098] The local loss function expression for the k-th iteration of the AE model is:

[0099]

[0100] in, Representing a subset The quantity value;

[0101] In the k-th iteration, the local AE model is updated as follows:

[0102]

[0103] in, This represents the AE model parameters for user i in the r-th round and k-th iteration. express gradient, Let η represent the local loss function of the AE model in the k-th iteration. a This represents the learning rate of the AE model;

[0104] The training data after e iterations of AE model training is input into the encoder network of the AE model to obtain the latent feature vector. The latent feature vectors are then input into the DDPM model for e iterations of training. The expression is:

[0105]

[0106] in, Let B represent the AE model parameters for user i in the r-th iteration. i Represents the training set;

[0107] The process of training the DDPM model includes:

[0108] In the k-th iteration, user i starts from the local training set. Randomly select a subset

[0109] Local loss function of DDPM model The expression is:

[0110]

[0111] in, Representing data loss function, Representing a subset One of the data points, This represents the DDPM model parameters for user i in the r-th round and k-th iteration;

[0112] In the k-th iteration, the local DDPM model is updated as follows:

[0113]

[0114] in, This represents the DDPM model parameters for user i in the r-th round and k-th iteration. express gradient, Let η represent the local loss function of the k-th iteration of the DDPM model. d This represents the learning rate of the DDPM model;

[0115] After completing e iterations of DDPM model training, the r-th round of local training ends, yielding the r-th round AE-DDPM model ω. r :

[0116] ω r ={ω r,a ,ω r,d}

[0117] Where, ω r,a ω r,dLet represent the parameters of the AE model and the DDPM model in the r-th round, respectively;

[0118] S203, Model Upload: The user will upload the updated AE-DDPM model ω from the rth round. r Uploaded from the local server to the base station;

[0119] S204, Model Aggregation: The base station calculates the weighted sum of the AE-DDPM models of all users within its coverage area to obtain a new global AE-DDPM model ω. r+1 The expression is:

[0120]

[0121] Where, d i d represents the local server data size of user i, d represents the total data size of all users within the base station coverage area, and η represents the learning rate of the aggregation model.

[0122] At this point, the r-th round of AE-DDPM model training is complete, and the base station has obtained a new global model;

[0123] S205, The new global AE-DDPM model ω r+1 Used for the next round of training, when the number of training rounds reaches a preset threshold R. max Training ends, and the final AE-DDPM model is obtained;

[0124] The pseudocode for the AE-DDPM model training algorithm based on federated learning is shown below:

[0125]

[0126] S3. Obtain the global predicted content popularity and cache the N most popular contents according to the base station's cache capacity;

[0127] The base station uses the global DDPM model ω d Perform a backdiffusion process and generate U spurious samples g. fake,u u = 1, 2, ..., U;

[0128] U fake samples g fake,u Input AE model decoder network to generate reconstructed fake samples The expression is:

[0129]

[0130] Where, ω a D represents the global AE model, D(·) represents the decoder of the AE model, and U represents the number of fake samples;

[0131] The reconstructed fake sample Instead of user data, a content score is calculated to predict content popularity at base stations; the calculation of the content score includes the following steps:

[0132] The reconstructed fake sample By adding the dimensions together, we can obtain a score for all content. The expression is:

[0133]

[0134] Where F represents the number of content types contained in the content library, and F represents the reconstructed fake sample. The dimension; It reflects the overall preferences of users within the base station's coverage area and does not expose the privacy of individual users. The higher the score, the more popular the content. Based on the base station's cache capacity, the N most popular pieces of content are cached.

[0135] like Figure 2 As shown, the base station cache hit rate of various methods is compared under different cache capacities. With the increase of cache capacity, the cache hit rate of various caching methods all improves. This is because a larger cache capacity allows the base station to store more content, making it more likely that users will obtain the requested content from the base station. The method proposed in this embodiment and CPCP outperform Thompson Sampling and N-greedy. This is because Thompson Sampling does not rely on learning-based content prediction, while N-greedy only caches the most frequently requested content and does not consider the potential features in the data. In addition, the method proposed in this embodiment outperforms CPCP because DDPM utilizes a progressive denoising generation process, a stable training objective, and a more comprehensive data distribution approximation capability, effectively overcoming the limitations of generative adversarial networks (GANs) in terms of training instability and pattern collapse. Oracle has the highest cache hit rate because it knows the user's future request content in advance, which is the highest value that the cache hit rate can achieve.

[0136] like Figure 3 As shown, the request content latency of various caching methods under different cache capacities is as follows: as the cache capacity increases, the request content latency of all methods decreases. This is because a larger cache capacity allows the base station to store more content, thereby increasing the likelihood that each user can directly obtain the required content from the base station, thus reducing the request latency. In addition, the request latency of the method provided in this embodiment is lower than that of other methods except Oracle. This is attributed to the fact that the method proposed in this embodiment has a higher cache hit rate, enabling more users to obtain content from the base station and reducing the content request latency.

[0137] like Figure 4 As shown, in the first nine training rounds, the cache hit rate gradually increases and the content request latency gradually decreases; this is because the base station gradually caches suitable popular content, and the model tends to converge around the ninth round.

[0138] like Figure 5 As shown, the hit rate of the caching method provided in this embodiment gradually increases as the number of users participating in the training increases; this is because more users provide more data and computing power, thus enabling more accurate prediction of popular content.

[0139] like Figure 6 The cache hit rate of different models changes with training rounds. The AE-DDPM model performs better than DDPM in terms of cache hit rate. This is because DDPM has difficulty learning an effective distribution from sparse user data, while AE-DDPM uses the AE model to extract latent feature vectors from user data for training DDPM, enabling DDPM to learn the distribution of user data better and thus improving performance.

[0140] By comparing different caching methods, the AE-DDPM model provided in this embodiment has the advantages of high cache hit rate, low request latency, improved model convergence trend and performance, and increased cache hit rate with an increased number of users.

[0141] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A federated learning-assisted edge caching method based on the AE-DDPM model, characterized in that, Includes the following steps: S1. Establish an edge computing network system model, including base stations, remote cloud servers, and users; S2. The AE-DDPM model based on federated learning is used for training; S201, Base station generates global initial AE-DDPM model The global initial AE-DDPM model Distribute to users for training; S202 and AE models are run locally. In each iteration of training, latent feature vectors are extracted from the user's training data. And used in the DDPM model Training iterations; The process of training the AE model includes: In the In this iteration, the user From local training set Randomly select a subset ; The subset training data points The data is input into an After Effects (AE) model, which then generates reconstructed output data points. In the In this iteration, the local AE model is updated as follows: in, Indicates the first Round In the next iteration, users AE model parameters, express gradient, Indicates the first The local loss function of the next iteration of the AE model. This represents the learning rate of the AE model; Will be completed The training data after each iteration of the AE model training is input into the encoder network of the AE model to obtain the latent feature vector. And input it into the DDPM model for processing. Training iterations; The process of training the DDPM model includes: In the In this iteration, the user From local training set Randomly select a subset ; In the In this iteration, the local DDPM model is updated as follows: in, Indicates the first Round In the next iteration, users DDPM model parameters, express gradient, Indicates the first The local loss function of the next iteration of the DDPM model. This represents the learning rate of the DDPM model; Finish The DDPM model iteration training ends. After rounds of local training, the first Wheel AE-DDPM model : in, , They represent the first Parameters of the in-wheel AE model and DDPM model; S203, The user will update the... Wheel AE-DDPM model Uploaded from the local server to the base station; S204. The base station calculates the weighted sum of the AE-DDPM models of all users within its coverage area to obtain a new global AE-DDPM model. ; S205, The new global AE-DDPM model Used for the next round of training, when the number of training rounds reaches a preset threshold. Training ends, and the final AE-DDPM model is obtained; S3. Obtain the global predicted content popularity, and cache the N most popular contents based on the base station's cache capacity. The predicted content popularity includes: Base stations use a global DDPM model Perform the reverse diffusion process and generate A fake sample ; Will The aforementioned fake samples Input AE model decoder network to generate reconstructed fake samples The expression is: in, Represents the global AE model. The decoder representing the AE model. Indicates the number of fake samples; The reconstructed fake sample Instead of user data, content scores are calculated to predict content popularity in base stations.

2. The federated learning-assisted edge caching method based on the AE-DDPM model according to claim 1, characterized in that, The data points generated by the AE model reconstruction The expression is: in, Indicates the first Round In the next iteration, users AE model parameters, , These represent the decoder and encoder of the AE model, respectively; The data points The loss function expression is: No. The local loss function expression for the next iteration of the AE model is: in, Representing a subset The quantity and size value.

3. The federated learning-assisted edge caching method based on the AE-DDPM model according to claim 2, characterized in that, The latent feature vector The expression is: in, Indicates the first User in round iteration AE model parameters, This represents the local training set.

4. The federated learning-assisted edge caching method based on the AE-DDPM model according to claim 3, characterized in that, The local loss function of the DDPM model The expression is: in, Representing data loss function, Representing a subset One of the data points, Indicates the first Round In the next iteration, users DDPM model parameters.

5. A federated learning-assisted edge caching method based on the AE-DDPM model according to claim 1, characterized in that, The new global AE-DDPM model The expression is: in, Indicates user The size of the local server data. This represents the total data size of all users within the base station's coverage area. This represents the learning rate of the aggregation model.

6. The federated learning-assisted edge caching method based on the AE-DDPM model according to claim 1, characterized in that, Calculating content scores involves the following steps: The reconstructed fake sample By adding the dimensions together, we can obtain a score for all content. The expression is: in, This indicates the number of content types contained in the content library. Indicates the reconstruction of fake samples The higher the score, the more popular the content.

Images