RL-Enhanced Resource Allocation Optimization for Efficient AIGC-Based Federated Learning at Edge

Abstract

Foundation models (FMs) are increasingly employed in federated learning (FL) at edge to support artificial intelligence generated content (AIGC) services. However, the large-scale parameterization and extensive pretraining requirements of FMs impose significant computational and communication burdens on resource-constrained edge networks, resulting in considerable delays in AIGC services. Current model lightweighting methods require additional training to compensate for model degradation. Furthermore, most bandwidth allocation methods ignore channel characteristics such as signal-to-noise ratio (SNR), packet loss rate (PLR), and transmission power of edge devices with weak connectivity, resulting in frequent retransmission and low transmission rates. In order to improve the efficiency of AIGC-based FL, both local training phase and model uploading phase for global aggregation are considered. We propose a lightweight training mechanism based on an optimized proximal policy optimization (PPO) algorithm, which adaptively adjust the number of heads in attention layer and neurons in the feedforward layer. Then we propose a comprehensive bandwidth resource allocation mechanism for efficient gradient transmission based on a gated recurrent unit (GRU)-enhanced asynchronous advantage actor-critic (A3C) algorithm, which considers SNR, PLR, and transmission power. Simulations demonstrate that the proposed methods achieve faster FL without model degradation compared with related approaches.