Fed-AGRU-Net: A Privacy-Preserving Federated Attention-GRU Framework for Reliable Emergency Message Dissemination in VANETs

Abstract

In Vehicular Ad Hoc Networks (VANETs), the topology is constantly changing, the channels are congested, the communication resources are limited, and privacy issues can be an obstacle that impede timely dissemination of safetycritical information. The current federated learning methods can support collaborative intelligence within retaining the vehicular data local. In addition, federated misbehaviour detection boosts the ability to detect unreliable communication behaviour, federated cyberattack detection enhances the security of the vehicular network and differential-privacy-based federated learning adds extra protection for distributed model updates. Nevertheless, most of them adopt conventional architectures or any stand-alone recurrent model and might not capture the complex temporal-feature dependency or the most safety-critical features for communications in different heterogeneous and non-IID vehicular environments. To overcome these drawbacks, a new approach Fed-AGRU-Net: A Privacy-Preserving Federated Attention-GRU Framework for Reliable Emergency Message Dissemination in VANETs is proposed. The proposed framework includes stacked Bidirectional Gated Recurrent Units (BiGRU), multi-head self-attention and a residual gated feature-recalibration mechanism, enabling the capture of temporal relationships and highlighting relevant dissemination features. Non-IID Dirichlet based client partitioning, weighted FedAvg, server-side update damping, differential-privacy based clipping with gaussian noise further promote the collaboration learning stability and privacy protection. Experimental results conducted on VANET Emergency Communication Dataset are shown to be 96.11% accurate, 96.33% precise, 95.38% recalled, 95.85% F1-score, and 0.1364 loss. The proposed model also shows improved overall prediction accuracy and decreased loss compared to FLEMD, Fed-GRU, and DP-FL, which proves the effectiveness of attentionenhanced recurrent learning and privacy-preserving federated optimization in predicting the delivery of emergency messages in VANET environments under different scenarios.