Federal learning auxiliary edge caching method based on AE-DDPM model

Abstract

The invention relates to a federal learning auxiliary edge caching method based on an AE-DDPM model, and the method comprises the following steps: building an edge computing network system model which comprises a base station, a remote cloud server and a user; an AE-DDPM model based on federated learning is adopted for training; obtaining global prediction content popularity, and caching the most popular N contents according to the cache capacity of the base station; according to the method, potential feature vectors of user data are extracted through the AE model, the data dimension and sparsity are reduced, data distribution is learned through the DDPM model, and high-quality content popularity prediction is generated; according to the method, the cache unit is deployed at the edge node, so that the user can quickly acquire the pre-cached content on the node, the cache hit rate is effectively improved, the time delay of acquiring the content by the user is reduced, the communication efficiency is remarkably improved, and meanwhile, the risk of privacy disclosure of the user is reduced.

Description

Technical Field

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

[0002] In recent years, with the rapid development of the mobile internet and the increasing popularity of smart devices, mobile data traffic has seen explosive growth, placing tremendous pressure on wireless networks. Edge caching technology has emerged in response to this situation. Edge caching deploys caching units at network edge nodes, such as base stations and access points, to pre-store content that users may be interested in. When users request this content, they can retrieve it directly from nearby edge nodes, eliminating the need to download it from distant data centers or the cloud. This reduces content delivery latency, alleviates network congestion, and improves user experience and overall system performance.

[0003] In order to effectively enable users to obtain content of interest from nearby base stations, base stations must predict popular content based on the preferences of users in their coverage area. Machine learning technology can extract the potential features of user data and predict preferred content. However, user personal data usually contains a large amount of privacy-sensitive information, and users usually refuse to share data directly with others. Traditional centralized machine learning methods require collecting user data to a central server for training, which not only has the potential to infringe user privacy, but also faces problems such as data leakage and high transmission costs. The emergence of 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. Federated learning avoids the risk of leaking user privacy information by sharing other information such as local model parameters rather than raw data, providing a reliable way to train models in scenarios where data is widely distributed and privacy-sensitive, such as mobile devices and IoT devices.

[0004] Although traditional federated learning methods can protect user privacy to a certain extent, their communication efficiency is low. Participants need to communicate frequently and the amount of data is extremely large, which leads to increased network congestion and latency. In addition, when processing sparse data, traditional federated methods have limited model performance and it is difficult to accurately extract effective features, which affects 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 of interest to users while protecting user privacy, and cache the predicted content on the base station, which can effectively reduce the delay for users to obtain content of interest.

[0006] To achieve the above objectives, the present 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, using the federated learning-based AE-DDPM model for training;

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

[0010] S202, the AE model performs e iterations of training locally to extract the potential feature vectors in the user training data And used for e-iteration training of DDPM model;

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

[0012] In the kth iteration, user i selects i Randomly select a subset

[0013] The subset The training data point z is input to the AE model, and the AE model generates the reconstructed output data point In the kth iteration, the local AE model is updated as:

[0014]

[0015] in, represents the AE model parameters of user i in the kth iteration of the rth round, express The gradient, represents the local loss function of the k-th iteration AE model, η a Represents the learning rate of the AE model;

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

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

[0018] In the kth iteration, user i selects Randomly select a subset

[0019] In the kth iteration, the local DDPM model is updated as:

[0020]

[0021] in, represents the DDPM model parameters of user i in the kth iteration of the rth round, express The gradient, represents the local loss function of the k-th iteration DDPM model, η d Represents the learning rate of the DDPM model;

[0022] Complete e DDPM model iterative training, end the rth round of local training, and obtain the rth round AE-DDPM model ω r :

[0023] ω r ={ω r,a ,ω r,d}

[0024] Among them, ω r,a 、ω r,d They represent the parameters of the AE model and DDPM model in the rth round respectively;

[0025] S203: The user updates the r-th round AE-DDPM model ω r Uploaded to the base station by the local server;

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

[0027] S205: transform the new global AE-DDPM model ω r+1 For the next round of training, when the number of training rounds reaches the preset threshold R max , end the training and obtain the final AE-DDPM model;

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

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

[0030]

[0031] in, represents the AE model parameters of user i in the kth iteration of the rth round, D(·) and E(·) represent the decoder and encoder of the AE model respectively;

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

[0033]

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

[0035]

[0036] in, Representation subset The quantity value.

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

[0038]

[0039] in, represents the AE model parameters of user i in the rth iteration, B i represents the training set.

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

[0041]

[0042] in, Representation data The loss function is Representation subset A data point in represents the DDPM model parameters of user i in the kth iteration of the rth round.

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

[0044]

[0045] Among them, d i represents the local server data size of user i, d represents the total data size of all users within the coverage of the base station, 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 the back diffusion process and generate U false samples g fake,u ;

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

[0049]

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

[0051] Reconstruct the false sample Instead of user data, content scores are calculated to predict content popularity in the base station.

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

[0053] Reconstruct the false sample By adding up the dimensions, we can get the scores of all the content. The expression is:

[0054]

[0055] Among them, F represents the number of content types contained in the content library, and F represents the reconstructed false samples. dimension; the higher the score, the more popular the content.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] The present invention extracts potential feature vectors of user data through the AE model, reduces data dimensionality and sparsity, and provides high-quality input for the DDPM model. DDPM generates the data distribution required for content popularity prediction through gradual denoising, improves prediction accuracy, and solves the limitations of traditional federated learning when processing sparse data. At the same time, the present invention can more accurately predict content popularity, enabling base stations to cache content that better meets user needs, significantly improving the cache hit rate, reducing the delay in users obtaining content, improving user experience and overall system performance, and further reducing the risk of user privacy leakage. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of the scenario of the present invention;

[0059] Figure 2 Graph showing how cache hit rates vary with cache capacity for different caching methods;

[0060] Figure 3Graph showing how content request latency varies with cache capacity for different caching methods;

[0061] Figure 4 This is a relationship diagram between cache hit rate and request content delay of the present invention;

[0062] Figure 5 This is a graph showing how the cache hit rate of the present invention changes with the number of users participating in training;

[0063] Figure 6 This is a graph showing how cache hit rates for different models of the present invention change with training rounds. DETAILED DESCRIPTION

[0064] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0065] Existing federated methods have difficulty in accurately extracting effective features when processing sparse data, which affects the accuracy of content preference prediction. The autoencoder model (AE) is an unsupervised learning neural network model, mainly composed of an encoder and a decoder: the encoder maps the input data to a low-dimensional latent space and extracts the feature vector of the data; the decoder reconstructs the original data based on the latent feature vector. Autoencoders have a wide range of applications in feature extraction, dimensionality reduction, image generation and other fields. They can learn efficient representations of data and provide better input features for machine learning tasks. The denoising difficult probabilistic model (DDPM) is a generative model based on the latent diffusion process. Its core idea is to gradually add noise to the data to gradually transform the data distribution into a noise distribution, and then train a neural network to learn the inverse denoising process to restore the original data from the noise. The DDPM model has demonstrated strong performance in tasks such as image generation, speech synthesis, and data interpolation, and can generate high-quality samples.

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

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

[0068] The base station and remote server are connected via a reliable backhaul link. All users are within the base station's coverage area, and each user i owns a smart device. The base station has limited storage capacity and can store up to N pieces of content simultaneously, while the remote cloud server caches all available content. If 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 higher content request latency.

[0069] The communication model used in this embodiment is as follows: multiple users communicate with the base station using Orthogonal Frequency Division Multiplexing (OFDM) in the sub-6 GHz frequency band. 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 represents the available communication bandwidth between the base station and user i, P s Indicates the transmit power of the base station, is the noise power at the receiver, g i represents the small-scale fading factor, which obeys an exponential distribution with a mean value of 1, h i represents the large-scale fading factor, including the effects of shadowing and path loss; shadowing 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, using the federated learning-based AE-DDPM model 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 the scheduling strategy Parameterized Markov chain, Gaussian noise is gradually added to the original data, so that the data distribution is gradually disturbed to random noise. The single-step diffusion process expression is:

[0077]

[0078] Among them, x t represents the data of diffusion step t, represents Gaussian distribution, β t represents the parameter controlling the noise level;

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

[0080] definition:

[0081]

[0082] Then we can get:

[0083]

[0084] Backward diffusion process: By training the neural network μ θ The noise ∈ of each prediction step θ , recover the original data distribution from the 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 the global initial AE-DDPM model ω 0 , the AE-DDPM model consists of AE and DDPM; ω r represents the global AE-DDPM parameters of the rth round of training. For subsequent rounds, the base station updates the global model at the end of the previous round of training. r,a 、ω r,d denote the parameters of AE and DDPM in the rth round, so 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 is trained locally for e iterations to extract the potential feature vectors in the user training data The extracted data is then used for e iterations of training of the DDPM model;

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

[0092] In the kth iteration, user i selects i Randomly select a subset

[0093] Subset The training data point z is input into the AE model, and the AE model generates the reconstructed output data point The expression is:

[0094]

[0095] in, represents the AE model parameters of user i in the kth iteration of the rth round, D(·) and E(·) represent the decoder and encoder of the AE model respectively;

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

[0097]

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

[0099]

[0100] in, Representation subset The quantity value of

[0101] In the kth iteration, the local AE model is updated as:

[0102]

[0103] in, represents the AE model parameters of user i in the kth iteration of the rth round, express The gradient, represents the local loss function of the k-th iteration AE model, η a Represents the learning rate of the AE model;

[0104] The training data after e times of AE model iteration training is input into the encoder network of the AE model to obtain the potential feature vector And input into the DDPM model for e iterations of training, the potential feature vector The expression is:

[0105]

[0106] in, represents the AE model parameters of user i in the rth iteration, B i represents the training set;

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

[0108] In the kth iteration, user i selects Randomly select a subset

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

[0110]

[0111] in, Representation data The loss function is Representation subset A data point in represents the DDPM model parameters of user i in the k-th iteration of the r-th round;

[0112] In the kth iteration, the local DDPM model is updated as:

[0113]

[0114] in, represents the DDPM model parameters of user i in the kth iteration of the rth round, express The gradient, represents the local loss function of the k-th iteration DDPM model, η d Represents the learning rate of the DDPM model;

[0115] Complete e DDPM model iterative training, end the rth round of local training, and obtain the rth round AE-DDPM model ω r :

[0116] ω r ={ω r,a ,ω r,d}

[0117] Among them, ω r,a 、ω r,dThey represent the parameters of the AE model and DDPM model in the rth round respectively;

[0118] S203, model upload: the user updates the rth round AE-DDPM model ω r Uploaded to the base station by the local server;

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

[0120]

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

[0122] At this point, the rth round of AE-DDPM model training is completed, and the base station obtains a new global model;

[0123] S205, the new global AE-DDPM model ω r+1 For the next round of training, when the number of training rounds reaches the preset threshold R max , end the training and obtain the final AE-DDPM model;

[0124] The pseudo code of the AE-DDPM model training algorithm based on federated learning is as follows:

[0125]

[0126] S3. Obtain global predicted content popularity and cache the N most popular contents based on the cache capacity of the base station;

[0127] The base station uses the global DDPM model ω d Perform the back diffusion process and generate U false samples g fake,u ,u=1,2,…,U;

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

[0129]

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

[0131] Reconstruct the false sample Instead of user data, a content score is calculated to predict the popularity of content in the base station; calculating the content score includes the following steps:

[0132] Reconstruct the false sample By adding up the dimensions, we can get the scores of all the content. The expression is:

[0133]

[0134] Among them, F represents the number of content types contained in the content library, and F represents the reconstructed false samples. Dimensions; It reflects the overall preference of users within the base station coverage area without exposing the privacy of individual users. The higher the score, the more popular the content. Based on the cache capacity of the base station, the N most popular contents are cached.

[0135] like Figure 2 As shown in the figure, the base station cache hit rates of various methods under different cache capacities are compared. As the cache capacity increases, the cache hit rates of various cache methods all improve. This is because a larger cache capacity enables 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 CPPPP outperform Thompson Sampling and N-greedy 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 CPPPP because DDPM utilizes a stepwise denoising generation process, a stable training objective, and a more comprehensive data distribution approximation capability to effectively overcome the limitations of generative adversarial networks (GANs) in terms of training instability and mode collapse. Oracle has the highest cache hit rate because it knows the user's future request content in advance, which is the highest cache hit rate that can be achieved.

[0136] like Figure 3 Figure 2 shows the request content delays of various caching methods at different cache capacities. As cache capacity increases, the request content delays of all methods decrease. This is because a larger cache capacity enables the base station to store more content, thereby increasing the likelihood that each user can obtain the required content directly from the base station, thereby reducing the request delay. Furthermore, the request delay of the method provided by this implementation is lower than that of other methods except Oracle. This is attributed to the higher cache hit rate of the method proposed by this implementation, which enables more users to obtain content from the base station and reduces the content request delay.

[0137] like Figure 4 As shown, in the first nine rounds of training, the cache hit rate gradually increases and the content request latency gradually decreases as the number of training rounds increases. This is because the base station gradually caches appropriate popular content, and the model tends to converge after about nine rounds.

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

[0139] like Figure 6 The figure shows how the cache hit rates of different models change with training rounds. The AE-DDPM model performs better than DDPM in terms of cache hit rate. Since DDPM has difficulty learning an effective distribution from sparse user data, AE-DDPM uses the AE model to extract latent feature vectors from user data for DDPM training, enabling DDPM to better learn the distribution of user data and thus improve 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, model convergence trend and performance improvement, and the increase in the number of users can improve the cache hit rate.

[0141] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as above in terms of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can, without departing from the scope of the technical solution of the present invention, make some changes or modifications to equivalent embodiments using the technical contents disclosed above. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A federated learning-assisted edge caching method based on the AE-DDPM model, characterized in that: The following steps are involved: S1. Establish an edge computing network system model, including base stations, remote cloud servers, and users; S2, using the federated learning-based AE-DDPM model for training; S201, the base station generates a global initial AE-DDPM model ω 0 , the global AE-DDPM model ω 0 Distribute to users for training; S202, the AE model performs e iterations of training locally to extract the potential feature vectors in the user training data And used for e-iteration training of DDPM model; The process of training the AE model includes: In the kth iteration, user i selects i Randomly select a subset The subset The training data point z is input to the AE model, and the AE model generates the reconstructed output data point In the kth iteration, the local AE model is updated as: in, represents the AE model parameters of user i in the kth iteration of the rth round, express The gradient, represents the local loss function of the k-th iteration AE model, η a Represents the learning rate of the AE model; The training data after e times of AE model iteration training is input into the encoder network of the AE model to obtain the potential feature vector And input it into the DDPM model for e iteration training; The process of training the DDPM model includes: In the kth iteration, user i selects Randomly select a subset In the kth iteration, the local DDPM model is updated as: in, represents the DDPM model parameters of user i in the kth iteration of the rth round, express The gradient, represents the local loss function of the k-th iteration DDPM model, η d Represents the learning rate of the DDPM model; Complete e DDPM model iterative training, end the rth round of local training, and obtain the rth round AE-DDPM model ω r : oh r ={ω r,a ,oh r,d } Among them, ω r,a 、ω r,d They represent the parameters of the AE model and DDPM model in the rth round respectively; S203: The user updates the r-th round AE-DDPM model ω r Uploaded to the base station by the local server; S204: The base station calculates the weighted sum of the AE-DDPM models of all users within the coverage area to obtain a new global AE-DDPM model ω r+1 ; S205: transform the new global AE-DDPM model ω r+1 For the next round of training, when the number of training rounds reaches the preset threshold R max , end the training and obtain the final AE-DDPM model; S3. Obtain global predicted content popularity and cache the most popular N contents based on the cache capacity of the base station.

2. The method for federated learning-assisted edge caching based on the AE-DDPM model according to claim 1, characterized in that: The AE model reconstructs the generated data points The expression is: in, represents the AE model parameters of user i in the kth iteration of the rth round, D(·) and E(·) represent the decoder and encoder of the AE model respectively; The loss function expression of the data point z is: The local loss function expression of the k-th iteration AE model is: in, Representation subset The quantity value.

3. The method for federated learning-assisted edge caching based on the AE-DDPM model according to claim 2, characterized in that: The latent feature vector The expression is: in, represents the AE model parameters of user i in the rth iteration, B i represents the training set.

4. The method for federated learning-assisted edge caching 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, Representation data The loss function is Representation subset A data point in represents the DDPM model parameters of user i in the kth iteration of the rth round.

5. The method for federated learning-assisted edge caching based on the AE-DDPM model according to claim 1, characterized in that: The new global AE-DDPM model ω r+1 The expression is: Among them, d i represents the local server data size of user i, d represents the total data size of all users within the coverage of the base station, and η represents the learning rate of the aggregation model.

6. The method for federated learning-assisted edge caching based on the AE-DDPM model according to claim 1, characterized in that: Predicting content popularity involves the following steps: The base station uses the global DDPM model ω d Perform the back diffusion process and generate U false samples g fake,u ; The U false samples g fake,u Input AE model decoder network to generate reconstructed fake samples The expression is: Among them, ω a represents the global AE model, D(·) represents the decoder of the AE model, and U represents the number of false samples; Reconstruct the false sample Instead of user data, content scores are calculated to predict content popularity in the base station.

7. The method for federated learning-assisted edge caching based on the AE-DDPM model according to claim 6, characterized in that: Calculating the content score involves the following steps: Reconstruct the false sample By adding up the dimensions, we can get the scores of all the content. The expression is: Among them, F represents the number of content types contained in the content library, and F represents the reconstructed false samples. dimension; the higher the score, the more popular the content.

Images