TPS: Trust-Aware Pruning for Byzantine Robustness Federated Learning in Real-Time Edge Systems

Federated learning is an essential contributor to future-generation digital ecosystems, but currently has two challenges: how to ensure communication efficiency and simultaneously provide Byzantine resilience. The current methods frequently focus on either one or another of these issues concentrating on a single either model compression to reduce bandwidth or strong aggregation to ensure security but does not focus on the interaction between the two. Trust-aware Pruning Strategy Trust -aware Pruning Strategy (TPS) is a framework that is presented in this paper and provides synergistic integration of compression- aware mechanisms with trust- based client evaluation. TPS attains communication efficiency and adversarial resilience, promoting the field of secure distributed learning on real-time edge settings. TPS proposes a decentralized structure where the client are able to engage in organized pruning whereby the pruning ratio of clients is variable and peer to peer validation is done to evaluate who to trust. The system uses a dynamic trust graph to model validation relationships and extract trust scores that form a difference between valid compression strategies and malicious manipulation. Trust-weighted averaging of thin slice model updates ensures model integrity and greatly reduces communication overhead, which is a highly important concern in real-time distributed application. Extensive testing on the MNIST proves that TPS can attain the test accuracy of about 89% by compression-based attacks with the adversarial participation rate of 20%. Moreover, it holds 80% accuracy in the worst-case situation of non-IID, which confirms its strength in a heterogeneous setting.