SecureGraphFL: A Privacy-Preserving and Attack-Resilient Federated Learning Framework for Traffic Prediction

Traffic prediction is a critical research topic with significant transportation management and planning implications. Graph Neural Networks (GNNs) have gained widespread use in traffic prediction due to their ability to capture relationships between different traffic sensors effectively. However, traditional centralized learning approaches struggle to meet the increasing demand for user data privacy. While federated learning (FL) addresses these privacy concerns, it introduces new challenges, particularly when one or more clients are compromised by malicious attacks or single-point failures, leading to significant degradation in global model performance. To address these issues, we propose SecureGraphFL, a federated learning framework based on graph convolutional networks (GCNs) and actor-critic networks that ensures security while maintaining high predictive accuracy. SecureGraphFL introduces three key innovations: (1) incorporating a parallel speed difference prediction task to enhance prediction performance, (2) utilizing a cosine similarity-based graph construction method to more effectively capture patterns from similar traffic sensors, and (3) designing an actor-critic network-based client selection mechanism to exclude malicious clients, ensuring the stability of the global model. Experiments on two open-source datasets, PEMS-D7 and METR-LA, demonstrate that in the absence of malicious or failed clients, SecureGraphFL matches state-of-the-art centralized methods and significantly outperforms traditional FL algorithms like FedAVG and FedATT. Moreover, in scenarios involving malicious clients, our client selection mechanism effectively mitigates their impact, maintaining performance levels comparable to those without any malicious activity. The code for this paper is available at: https://github.com/cry-C/SecureGraphFL.