Explainable Graph Attention-Driven Fairness Manipulation for Federated Learning in EdgeIoT

This paper proposes an innovative adversarial architecture based on Explainable Graph AtTention-embedded autoEncoder (E-GATE), specifically designed to execute fairness manipulation that introduce biasing model updates into the federated learning in edge-based Internet of Things (EdgeIoT). E-GATE aims to generate biasing model updates by maximizing the minimum Kullback-Leibler (KL) divergence between a device’s local model update and the global model. The E-GATE is trained with attention coefficients to obtain the hidden representations of each data feature in the explainable graph. Additionally, the graph autoencoder is incorporated within the E-GATE architecture to manipulatively reconstruct the correlations among model updates. This approach maximizes the reconstruction loss while keeping the biasing model updates undetected. The E-GATE attack is implemented using PyTorch, and experimental results demonstrate that it successfully increases the minimum KL divergence of benign model updates by $\mathbf{7 0. 2 \%}$, effectively evading detection by existing defense mechanisms.