Security-Aware Federated Graph Neural Network with Attention-Based Aggregation and Function-Hiding Encryption

Network security and privacy protection have encountered significant challenges in federated learning, particularly due to the risks of node attacks and data poisoning from participating nodes. Conventional methods often face difficulties in balancing data privacy and model performance. This paper introduces a Security-Aware Federated Graph Neural Network (Fed-GNN), designed to improve both the security and performance of federated learning systems. Fed-GNN introduces an attention-based aggregation algorithm on the server side, which dynamically evaluates the contribution of client models to the global model, thereby improving model robustness and adaptability. In addition, a function-hiding multi-input function encryption technique is employed to protect privacy during the transmission of model parameters. Experimental results demonstrate that Fed-GNN achieves F1 scores of 79.8% and 83.4% on the NYT and Web NLG datasets, respectively, outperforming baseline methods such as WDec and Span-RG. The F1 score on the NYT dataset shows a 4.1% improvement over Span-RG and a 37.8% improvement over Novel Tagging. In the overlapping triple extraction task, Fed-GNN enhances detection F1 scores for normal entity relations and multi-entity overlaps by 1.8% and 4.8%, respectively. In terms of privacy preservation, Fed-GNN achieves an accuracy of 97.86% under a privacy budget range of 36-45, even with 20 participating clients, surpassing several benchmark methods. Moreover, Fed-GNN improves detection capabilities for rare attack types such as R2L attacks, with a detection convergence speed approximately three times faster and a significant improvement in the final detection rate. These results demonstrate the effectiveness of Fed-GNN in enhancing both data security and model performance.