A semi-distributed federated learning collaborative optimization cache deployment system and method
By dividing the edge caching network into AP alliances and selecting representative APs for model training and caching decisions, the low utilization rate and privacy issues of edge caching resources are solved, achieving efficient caching optimization and model training, and adapting to the dynamic changes in user requests.
Patent Information
- Application Number
- CN202211119460.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-14
- Publication Date
- 2026-05-15
- Estimated Expiration
- 2042-09-14
AI Technical Summary
Existing federated learning frameworks cannot effectively optimize the resource utilization of edge cache points, and they also have privacy issues and insufficient model training accuracy, and cannot adapt to the randomness of user requests and geographical differences.
A semi-distributed federated learning collaborative optimization cache deployment system is adopted, which divides APs into different federations, selects representative APs through the intelligent gateway layer SG to perform model training and cache decision-making, and uses cross-media data feature similarity measurement for clustering to achieve distributed collaborative learning and cache optimization.
It improves cache hit rate and model training accuracy, protects user privacy, reduces communication costs and latency, and adapts to dynamic changes in user preferences.
Smart Images
Figure CN115665808B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to a semi-distributed federated learning collaborative optimization cache deployment system and method. Background Technology
[0002] The proliferation of smart mobile devices and the increasing use of various application types in recent years have led to an explosive growth in multimedia data such as text, images, videos, audio, and 3D models at the network edge, placing immense pressure on the fronthaul link. To cope with the massive content distribution and increasing demand for multimedia resources, and to support large-scale content delivery, simply increasing network capacity is insufficient. As an emerging technology addressing these challenges, edge caching is considered a prime example of alleviating unprecedented backhaul pressure and improving user QoE. It achieves low-latency, high-speed content delivery services by prefetching and storing a portion of popular multimedia content at the network edge. Given the high penetration and large scale of radio access, various access points (such as small cellular base stations or wireless access points supporting caching) with different radio access technologies (RATs), coverage areas, and backhaul capabilities become ideal carriers for edge caching. High deployment density allows APs to provide large coverage areas and high-quality wireless links, while edge caching alleviates fronthaul traffic bottlenecks and reduces service request latency by bringing content closer to users. Therefore, the system cost of high-speed fronthaul links can be controlled, improving the user experience.
[0003] However, the storage capacity of these access points is always limited, making it impossible to meet the needs of massive network scales. Therefore, optimizing the cache deployment of access points that initiate edge caching to maximize cache utilization becomes crucial. However, this is a rather challenging problem for the following reasons: In the real world, a considerable number of access points (APs) are deployed at the network edge, and APs in different geographical locations have different characteristics and play different roles. Therefore, a globally unique caching strategy for all APs seems too simplistic and cannot optimize the utilization of AP cache resources. Furthermore, due to the randomness of user requests, the popularity of content within the AP coverage area dynamically changes in the spatiotemporal dimension. To address this challenge, many recent works have proposed learning-based methods that centrally collect user data for training, but they bring a significant problem: users may distrust central servers and therefore hesitate to upload their private data. Moreover, a large amount of multimedia information such as text, images, and videos from different platforms and sources, while collectively depicting the same or related themes, presents complex, multi-layered semantic relationships, which traditional analysis and processing methods struggle to effectively utilize and learn from.
[0004] Even disregarding privacy concerns, traditional federated learning (FL) frameworks implicitly assume that a single model can simultaneously fit the data generation distributions of all edge endpoints in order to achieve joint training on large-scale distributed data. However, this is not the case in reality. Suppose multiple access points (APs) are attempting to jointly train a content popularity prediction model to predict future popular content files. In reality, due to the random nature of user requests, different APs may have different perceptions of the popularity of certain content. A single model can never accurately predict the popularity trend of a particular content file across all APs simultaneously. Furthermore, user request statistics within an AP's coverage area can vary significantly due to factors such as the AP's geographical location and preferences at different times. For example, APs located in busy areas typically display different user request statistics than APs in remote areas.
[0005] In all the above cases, the ordinary federated learning (FL) framework, in which all APs are treated equally and learn only a single global model, cannot achieve this goal. Summary of the Invention
[0006] To address the problems in the background technology, this invention proposes a semi-distributed federated learning collaborative optimization cache deployment system and method, the specific scheme of which is as follows:
[0007] A semi-distributed federated learning collaborative optimization cache deployment system is an edge caching network architecture consisting of a smart gateway layer (SG) and densely deployed wireless access point (AP) layers. The AP layers include, in sequence, a user request collection module, a request feature extraction module, a collaborative learning module, a local model training module, and a caching module. The user request collection module collects content request information from end users within the coverage area of the AP layers. The request feature extraction module extracts different media features containing fine-grained contextual information from the user request content and sends the information corresponding to different media features to the request feature extraction module of the smart gateway layer (SG). After the smart gateway layer (SG) determines the AP alliance and the representative AP, the collaborative learning module allows other non-representative APs within the alliance to share their locally collected data with the representative AP. In each communication round, the representative AP is responsible for training the local model and sends the updated local model parameters back to the smart gateway layer (SG) as feedback. The caching module prefetches specific media content into the local cache based on the caching decisions generated by the smart gateway layer (SG).
[0008] Specifically, the intelligent gateway layer SG includes a federation decision module, a global model aggregation module, and a cache decision module, set sequentially. The federation decision module divides all APs into different AP federations. The global model aggregation module generates a global model by calculating the weighted average of all received local models. The cache decision module generates a cache decision based on the content popularity prediction of the global model of deep learning file popularity prediction in the federated learning architecture.
[0009] Specifically, the alliance decision module collects user request data feature information sent by the wireless access point layer APs. Based on the hierarchical network architecture composed of smart gateways and ultra-densely deployed APs, it uses a clustering algorithm based on cross-media data feature similarity measurement to divide all APs into different AP alliances and determines the representative AP according to the principle of minimizing average transmission.
[0010] Based on the above method for a semi-distributed federated learning collaborative optimization cache deployment system, the process steps on the SG side of the intelligent gateway layer are as follows:
[0011] SA1, the alliance decision module of the intelligent gateway layer SG receives feature data requested by all wireless access point layer APs under the network domain;
[0012] SA2 and the Smart Gateway layer (SG) divide all APs into several distinct subgroups based on the request content of each wireless access point (AP) using a clustering algorithm that measures cross-media data feature similarity. Each subgroup forms an AP alliance.
[0013] SA3. Determine whether to start a new round of communication. If yes, proceed to step SA4; otherwise, proceed to step SA6.
[0014] SA4, the Smart Gateway Layer (SG) selects a representative AP within each AP consortium to send the global model based on the principle of minimizing the average transmission distance.
[0015] SA5 and the Smart Gateway layer (SG) transmit the global model to the representative AP in the wireless access point layer. The representative AP then transmits the global model to the corresponding AP alliance for training to obtain an updated local model. The global model aggregation module then aggregates all the local models obtained from the representative AP to form the updated global model.
[0016] SA6. After the communication round is completed, determine whether convergence has occurred. If convergence has not occurred, return to step SA3. If convergence has occurred, the cache decision module of the smart gateway layer SG determines to obtain the cache decision.
[0017] Based on the above-mentioned method for a semi-distributed federated learning collaborative optimization cache deployment system, the process on the AP side of the wireless access point layer is as follows:
[0018] SB1, After the user request collection module of the wireless access point layer AP collects the terminal user request data, the request feature extraction module extracts the user request data features.
[0019] SB2. Determine whether the wireless access point (AP) is a representative AP based on the SG alliance decision module of the smart gateway layer. If it is, the representative AP downloads the global model from the SG of the smart gateway layer. If not, proceed to step SB3.
[0020] SB3 represents the AP in collecting and storing local data from other APs within the alliance using the collaborative learning module, and in training the local model using the user request feature data extracted from all APs within the alliance through the local model training module. Then, it represents the AP in uploading the local model to the smart gateway layer SG.
[0021] After SB4 completes the communication round, it is determined whether convergence has occurred. When convergence occurs, it means that the AP receives the caching decision from the Smart Gateway Layer SG, and the caching module caches the content of the federated AP.
[0022] A method for a semi-distributed federated learning collaborative optimization cache deployment system includes the following steps:
[0023] S1. The wireless access point (AP) layer collects terminal user request data and then extracts the user request data features; it then sends the feature data of all wireless access point (AP) requests to the smart gateway (SG) layer.
[0024] S2, the Smart Gateway Layer (SG) divides all APs into several distinct subgroups based on the request content of each wireless access point (AP) using a clustering algorithm based on cross-media data feature similarity measurement. Each subgroup forms an AP alliance. A representative AP is selected within each AP alliance based on the principle of minimizing the average transmission distance.
[0025] S3. Determine whether to start a new round of communication. Based on the SG alliance decision module of the smart gateway layer, determine whether it represents the AP. If it does, the AP downloads the global model from the smart gateway layer SG and proceeds to step S5. If not, proceed to step S4.
[0026] S4 represents the AP using the collaborative learning module to collect and store local data from other APs within the alliance, and using the local model training module to train the local model using user request feature data extracted from all APs within the alliance. The AP represents the AP uploading all local models to the smart gateway layer SG, and the smart gateway layer SG aggregates them through the global model aggregation module to form an updated global model.
[0027] S5. After the communication round is completed, determine whether convergence has occurred. If convergence has not occurred, return to step S3. If convergence has occurred, the smart gateway layer SG determines to obtain the cache decision.
[0028] S6. The wireless access point (AP) receives the caching decision from the smart gateway (SG) and performs content caching for the alliance AP.
[0029] The beneficial effects of this invention are as follows:
[0030] (1) This application presents an edge caching architecture consisting of densely deployed fixed access points (APs) and intelligent gateways (SGs) to provide services to multiple fixed or low-mobility users. In the proposed architecture, we assume that the SG acts as a central coordinator to provide global services, such as caching popular content across the entire region, monitoring all user-generated requests in the local APs, centrally scheduling storage resources, and being responsible for executing cache placement decisions.
[0031] (2) To achieve joint training of large-scale distributed data and address the sensitive personal information of various end users, this application first proposes the concept of federated APs, dividing all APs into subgroups with the greatest differences, thus forming a novel semi-distributed learning architecture. Based on the above federation, distributed collaborative learning between APs is introduced into the local update process. Low-power communication between APs within the federation is used for collaboration and information exchange to improve model training performance and alleviate the difficulty of a single AP collecting enough massive feature data for model learning. Furthermore, we select representative APs to train the model. Since only one AP in the federation needs to upload its local model to the SG during global aggregation, unlike the traditional federated learning architecture where most devices need to upload their local models, this reduces the cost of resource-intensive uplink communication in traditional methods relative to star topologies, as well as the impact of unreliable communication on model accuracy. At the same time, it improves the scalability of the fully decentralized architecture through local low-power D2D communication.
[0032] (3) This application proposes a collaborative caching scheme based on federated learning under a semi-distributed architecture. By considering the dynamic characteristics of user preferences under the coverage of AP, the edge caching problem is studied. The smart gateway can personalize the storage of AP according to the available data to maximize the overall cache hit rate. Shared local data can increase the accuracy of model training and improve the accuracy of file popularity prediction.
[0033] (4) The caching decision module utilizes a deep learning-based global model for predicting file popularity. It uses user request feature data extracted from APs in the network domain to predict content popularity and generate caching decisions. Since all APs in the network domain collect user request data and capture feature information locally, and select representative APs to train models locally, only sending model parameters back to the SG to build an updated global model, there is no need to upload the user request data and feature information collected by all APs to the SG, thus protecting user privacy.
[0034] (5) Multiple APs within the consortium adopt a distributed collaborative learning mechanism to capture features from cross-media information requested by all AP users within the consortium and introduce them into a deep learning-based file popularity prediction model. Since APs within the consortium have the same preferences, the shared local data can increase the accuracy of model training and improve the accuracy of file popularity prediction. Attached Figure Description
[0035] Figure 1 This is a structural diagram of the Smart Gateway Layer (SG).
[0036] Figure 2 This is a structural diagram of the wireless access point (AP) layer.
[0037] Figure 3 This is the flowchart for the SG side of the smart gateway layer.
[0038] Figure 4 This is a flowchart for the AP side of the wireless access point layer. Detailed Implementation
[0039] like Figure 1-2 As shown, the semi-distributed federated learning collaborative optimization cache deployment system based on smart gateways is an edge cache network architecture consisting of a smart gateway layer (SG) and a densely deployed wireless access point layer (AP).
[0040] The Smart Gateway (SG) layer comprises a federation decision module, a global model aggregation module, and a cache decision module, arranged sequentially. The federation decision module collects user request data feature information from the Access Points (APs) at the wireless access point (AP) layer. Based on a hierarchical network architecture consisting of smart gateways and densely deployed APs, it uses a clustering algorithm based on cross-media data feature similarity measurement to divide all APs into different AP federations and determines representative APs based on the principle of minimizing average transmission. The global model aggregation module generates a global model by calculating the weighted sum of all received local models, which is used in the next communication round. The cache decision module uses a converged global model based on deep learning in a federated learning architecture to predict content popularity using data from APs in the network domain, generating cache decisions.
[0041] The wireless access point layer includes, in sequence, a user request collection module, a request feature extraction module, a collaborative learning module, a local model training module, and a caching module. The user request collection module collects content request information from terminal users within the coverage area of the wireless access point (AP). The request feature extraction module extracts different media features containing fine-grained contextual information from the user request content and sends the information corresponding to different media features to the smart gateway layer (SG). After the smart gateway layer (SG) determines the AP alliance and the representative AP, the collaborative learning module allows other non-representative APs within the alliance to share their locally collected data with the representative AP, improving the accuracy of its model training. In each communication round, the representative AP is responsible for training the local model and sends updates to the local model parameters back to the smart gateway layer (SG) as feedback. The caching module prefetches specific media content into the local cache based on the caching decisions generated by the smart gateway layer (SG).
[0042] This invention is based on an edge caching network architecture composed of smart gateways and ultra-densely deployed APs. Based on this layered architecture, the concept of a consortium is introduced. Due to significant differences in the geographical location and time-of-day preferences of different APs, we introduce a clustering algorithm based on cross-media data feature similarity measurement to divide all APs into the most dissimilar subgroups. Specifically, we perform clustering by analyzing the content features of user requests received by different APs. Although the multimedia objects requested by users, such as images, audio, and text, have different forms, there are many semantic relationships between them. After extracting different media features containing fine-grained contextual information, the aim is to learn a common shared space. By explicitly projecting data of different media types into this space, we can directly use general distance metrics (such as Euclidean distance and cosine distance) to measure the similarity between heterogeneous data objects and achieve clustering, thus constructing consortium APs. Then, based on the principle of minimizing average transmission distance, representative APs within the consortium are selected. The selected representative APs download the global model from the SG and introduce a distributed collaborative learning mechanism during local model training, accepting local data shared by other APs within the consortium to improve the accuracy of model training. In each communication round, the AP sends the updated model parameters back to the SG as feedback. After the AP uploads the model, the SG generates a new global model by calculating a weighted sum of all received local models. Finally, the deep learning-based file popularity prediction model in the SG uses data from the APs in the network domain to predict content popularity and prefetches the prediction results into the cache.
[0043] like Figure 3 As shown, based on the above structure, the process steps on the SG side of the smart gateway layer are as follows:
[0044] SA1, the alliance decision module of the intelligent gateway layer SG receives feature data requested by all wireless access point layer APs under the network domain;
[0045] SA2 and the Smart Gateway layer (SG) divide all APs into several distinct subgroups based on the request content of each wireless access point (AP) using a clustering algorithm that measures cross-media data feature similarity. Each subgroup forms an AP alliance.
[0046] SA3. Determine whether to start a new round of communication. If yes, proceed to step SA4; otherwise, proceed to step SA6.
[0047] SA4, the Smart Gateway Layer (SG) selects a representative AP within each AP consortium to send the global model based on the principle of minimizing the average transmission distance.
[0048] SA5 and the Smart Gateway layer (SG) transmit the global model to the representative AP in the wireless access point layer. The representative AP then transmits the global model to the corresponding AP alliance for training to obtain an updated local model. The global model aggregation module then aggregates all the local models obtained from the representative AP to form the updated global model.
[0049] SA6. After the communication round is completed, determine whether convergence has occurred. If convergence has not occurred, return to step SA3. If convergence has occurred, the cache decision module of the smart gateway layer SG determines to obtain the cache decision.
[0050] like Figure 4 As shown, the process on the AP side of the wireless access point layer is as follows:
[0051] SB1, After the user request collection module of the wireless access point layer AP collects the terminal user request data, the request feature extraction module extracts the user request data features.
[0052] SB2. Determine whether the wireless access point (AP) is a representative AP based on the SG alliance decision module of the smart gateway layer. If it is, the representative AP downloads the global model from the SG of the smart gateway layer. If not, proceed to step SB3.
[0053] SB3 represents the AP in collecting and storing local data from other APs within the alliance using the collaborative learning module, and in training the local model using the user request feature data extracted from all APs within the alliance through the local model training module. Then, it represents the AP in uploading the local model to the smart gateway layer SG.
[0054] After SB4 completes the communication round, it is determined whether convergence has occurred. When convergence occurs, it means that the AP receives the caching decision from the Smart Gateway Layer SG, and the caching module caches the content of the federated AP.
[0055] From a two-sided perspective, the method for semi-distributed federated learning to collaboratively optimize cache deployment systems includes the following steps:
[0056] S1. The wireless access point (AP) layer collects terminal user request data and then extracts the user request data features; it then sends the feature data of all wireless access point (AP) requests to the smart gateway (SG) layer.
[0057] S2, the Smart Gateway Layer (SG) divides all APs into several distinct subgroups based on the request content of each wireless access point (AP) using a clustering algorithm based on cross-media data feature similarity measurement. Each subgroup forms an AP alliance. A representative AP is selected within each AP alliance based on the principle of minimizing the average transmission distance.
[0058] S3. Determine whether to start a new round of communication. Based on the SG alliance decision module of the smart gateway layer, determine whether it represents the AP. If it does, the AP downloads the global model from the smart gateway layer SG and proceeds to step S5. If not, proceed to step S4.
[0059] S4 represents the AP using the collaborative learning module to collect and store local data from other APs within the alliance, and using the local model training module to train the local model using user request feature data extracted from all APs within the alliance. The AP represents the AP uploading all local models to the smart gateway layer SG, and the smart gateway layer SG aggregates them through the global model aggregation module to form an updated global model.
[0060] S5. After the communication round is completed, determine whether convergence has occurred. If convergence has not occurred, return to step S3. If convergence has occurred, the smart gateway layer SG determines to obtain the cache decision.
[0061] S6. The wireless access point (AP) receives the caching decision from the smart gateway (SG) and performs content caching for the alliance AP.
[0062] A computer device includes a processor and a storage medium, wherein a computer program is stored on the storage medium, and the processor reads from the storage medium and runs the computer program to execute the method described above for a semi-distributed federated learning collaborative optimization cache deployment system.
[0063] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A semi-distributed federated learning collaborative optimization cache deployment system, characterized in that, The architecture consists of an intelligent gateway layer (SG) and a densely deployed wireless access point layer (AP); the wireless access point layer (AP) includes, in sequence, a user request collection module, a request feature extraction module, a collaborative learning module, a local model training module, and a caching module. The user request collection module is used to collect content request information from terminal users within the coverage area of the wireless access point (AP) layer. The request feature extraction module is used to extract different media features containing fine-grained context information from the user request content, and send the information corresponding to the different media features to the request feature extraction module of the smart gateway layer SG. After the collaborative learning module determines the AP alliance and representative AP at the smart gateway layer SG, other non-representative APs in the alliance will share the data information collected locally with the representative AP. In each communication round, the local model training module is responsible for training the local model on behalf of the AP, and sends the update of the local model parameters back to the smart gateway layer SG as feedback; the caching module prefetches specific media content into the local cache according to the caching decision generated by the smart gateway layer SG. The system performs the following steps; S1. The wireless access point (AP) layer collects terminal user request data and then extracts the characteristics of the user request data. Send the characteristic data of all wireless access point (AP) requests to the smart gateway (SG) layer. S2, the Smart Gateway Layer (SG) divides all APs into several distinct subgroups based on the request content of each wireless access point (AP) using a clustering algorithm based on cross-media data feature similarity measurement. Each subgroup forms an AP alliance. A representative AP is selected within each AP alliance based on the principle of minimizing the average transmission distance. S3. Determine whether to start a new round of communication. Based on the SG alliance decision module of the smart gateway layer, determine whether it represents the AP. If it does, the AP downloads the global model from the smart gateway layer SG and proceeds to step S5. If not, proceed to step S4. S4 represents the AP using the collaborative learning module to collect and store local data from other APs within the alliance, and using the local model training module to train the local model using user request feature data extracted from all APs within the alliance. The AP represents the AP uploading all local models to the smart gateway layer SG, and the smart gateway layer SG aggregates them through the global model aggregation module to form an updated global model. S5. After the communication round is completed, determine whether convergence has occurred. If convergence has not occurred, return to step S3. If convergence has occurred, the smart gateway layer SG determines to obtain the cache decision. S6. The wireless access point (AP) receives the caching decision from the smart gateway (SG) and performs content caching for the alliance AP.
2. The semi-distributed federated learning collaborative optimization cache deployment system according to claim 1, characterized in that, The intelligent gateway layer SG includes a federation decision module, a global model aggregation module, and a cache decision module, which are set sequentially. The federation decision module divides all APs into different AP federations. The global model aggregation module generates a global model by calculating the weighted sum of all received local models. The cache decision module generates cache decisions based on the content popularity prediction of the global model.
3. The semi-distributed federated learning collaborative optimization cache deployment system according to claim 2, characterized in that, The alliance decision module collects user request data feature information sent by the wireless access point layer APs. Based on the hierarchical network architecture composed of smart gateways and ultra-densely deployed APs, it uses a clustering algorithm based on cross-media data feature similarity measurement to divide all APs into different AP alliances and determines the representative AP according to the principle of minimizing average transmission.
4. A semi-distributed federated learning collaborative optimization cache deployment method, characterized in that, The process steps on the SG side of the smart gateway layer, using the system of any one of claims 2-3, are as follows: SA1, the alliance decision module of the intelligent gateway layer SG receives feature data requested by all wireless access point layer APs under the network domain; SA2 and the Smart Gateway layer (SG) divide all APs into several distinct subgroups based on the request content of each wireless access point (AP) using a clustering algorithm that measures cross-media data feature similarity. Each subgroup forms an AP alliance. SA3. Determine whether to start a new round of communication. If yes, proceed to step SA4; otherwise, proceed to step SA6. SA4, the Smart Gateway Layer (SG) selects a representative AP within each AP consortium to send the global model based on the principle of minimizing the average transmission distance; SA5, the Smart Gateway layer SG transmits the global model to the representative AP in the wireless access point layer. The representative AP then transmits the global model to the corresponding AP alliance for training to obtain an updated local model. Then the global model aggregation module aggregates all the local models representing AP to form an updated global model; SA6. After the communication round is completed, determine whether convergence has occurred. If convergence has not occurred, return to step SA3. If convergence has occurred, the cache decision module of the smart gateway layer SG determines to obtain the cache decision.
5. The semi-distributed federated learning collaborative optimization cache deployment method according to claim 4, characterized in that, The process on the AP side of the wireless access point layer is as follows: SB1, After the user request collection module of the wireless access point layer AP collects the terminal user request data, the request feature extraction module extracts the user request data features. SB2. Determine whether the wireless access point (AP) is a representative AP based on the SG alliance decision module of the smart gateway layer. If it is, the representative AP downloads the global model from the SG of the smart gateway layer. If not, proceed to step SB3. SB3 represents the AP in collecting and storing local data from other APs within the alliance using the collaborative learning module, and in training the local model using the user request feature data extracted from all APs within the alliance through the local model training module. Then, it represents the AP in uploading the local model to the smart gateway layer SG. After SB4 completes the communication round, it is determined whether convergence has occurred. When convergence occurs, it means that the AP receives the caching decision from the Smart Gateway Layer SG, and the caching module caches the content of the federated AP.
6. A computer device, characterized in that, It includes a processor and a storage medium, on which a computer program is stored, and the processor reads from the storage medium and runs the computer program to execute a semi-distributed federated learning collaborative optimization cache deployment method as described in any one of claims 4-5.