Scholarlink

<p>The extensive adoption of smart vehicles and resource-hungry infotainment applications driven by Internet of Vehicles (IoV) technologies has led to the emergence of Vehicular Edge Computing (VEC) architectures. VEC focuses on efficiently processing data and tasks generated by vehicles close to their source, addressing latency and bandwidth constraints. However, intelligently caching content on edge servers for optimal performance while protecting consumer privacy introduces significant challenges. This paper proposes a hybrid Federated Learning and Deep Reinforcement Learning framework to enhance cache performance and preserve user privacy in AI-enabled VEC networks for consumer electronics. Long Short-Term Memory (LSTM) recurrent neural networks extract informative spatio-temporal features from real-world vehicle trace SUMO datasets, capturing mobility patterns, network traffic profiles, and privacy threats. These features model the VEC environment, including vehicle movements, content popularity dynamics, and privacy risks. Federated Learning enables collaborative model training across distributed edge nodes without sharing raw data, ensuring privacy preservation and reducing communication overhead. A Deep Q Network (DQN) agent, trained locally at each node and coordinated via federated aggregation, dynamically optimizes edge cache replacement policies. The framework balances objectives such as cache hit ratios, byte hit rates, access latency, and privacy goals, including location ambiguity and access pattern hiding. Comprehensive evaluations on vehicle traces confirm the proposed Federated LSTM-DQN model’s effectiveness in improving caching efficiency and privacy. The model achieves a training reward of 97.2%, validation reward of 96.88%, and average reward of 97.08% with low loss values indicating strong convergence and generalization. Privacy metrics also demonstrate reliable user privacy preservation alongside optimized caching.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Federated learning-based intelligent cache management for AI-enabled vehicular edge computing: optimizing efficiency and privacy using reinforcement learning

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Sohail Jabbar,
  • Muhammad Asif Habib,
  • Umar Raza,
  • Muhammad Farhan,
  • Ghufran Ahmed,
  • Muhammad Al-Abdullah

摘要

The extensive adoption of smart vehicles and resource-hungry infotainment applications driven by Internet of Vehicles (IoV) technologies has led to the emergence of Vehicular Edge Computing (VEC) architectures. VEC focuses on efficiently processing data and tasks generated by vehicles close to their source, addressing latency and bandwidth constraints. However, intelligently caching content on edge servers for optimal performance while protecting consumer privacy introduces significant challenges. This paper proposes a hybrid Federated Learning and Deep Reinforcement Learning framework to enhance cache performance and preserve user privacy in AI-enabled VEC networks for consumer electronics. Long Short-Term Memory (LSTM) recurrent neural networks extract informative spatio-temporal features from real-world vehicle trace SUMO datasets, capturing mobility patterns, network traffic profiles, and privacy threats. These features model the VEC environment, including vehicle movements, content popularity dynamics, and privacy risks. Federated Learning enables collaborative model training across distributed edge nodes without sharing raw data, ensuring privacy preservation and reducing communication overhead. A Deep Q Network (DQN) agent, trained locally at each node and coordinated via federated aggregation, dynamically optimizes edge cache replacement policies. The framework balances objectives such as cache hit ratios, byte hit rates, access latency, and privacy goals, including location ambiguity and access pattern hiding. Comprehensive evaluations on vehicle traces confirm the proposed Federated LSTM-DQN model’s effectiveness in improving caching efficiency and privacy. The model achieves a training reward of 97.2%, validation reward of 96.88%, and average reward of 97.08% with low loss values indicating strong convergence and generalization. Privacy metrics also demonstrate reliable user privacy preservation alongside optimized caching.