Detail

Toward Resilient CyberEdge Intelligence: Variational Graph Models and Data-Agnostic Poisoning in Federated Learning

In CyberEdge networks, where federated learning (FL) orchestrates intelligence across privacy-preserving edge devices, emerging model poisoning (MP) poses a critical threat to system resilience. This talk focuses on a new type of the data-untethered attack, where an adversarial variational graph autoencoder (VGAE) constructs malicious local model updates from benign updates, bypassing access to private training data. By extracting and regenerating high-order graph structural correlations among benign client models, the new VGAE-MP attack produces stealthy and effective poisoning that evades conventional detection and leads to a progressive degradation of global model accuracy. To counter such graph-driven, semantically aligned threats, we further present a visual explanation-based defense framework that transforms local updates into visual heat-map representations via Gradient-weighted Class Activation Mapping (GradCAM) and can enhance their separability through autoencoder-based feature projection. This approach exposes hidden discrepancies between benign and malicious updates that Euclidean distance/Cosine similarity-based defenses fail to capture. In addition, we discuss the escalating vulnerability of CyberEdge FL to advanced graph-based attacks as well as promising pathways toward more resilient and trustworthy distributed intelligence.