Conflict Aware Trust-Based Aggregation for Poisoning Resistant Federated Medical Image Classification

Introducing a lightweight conflict-aware trust-based aggregation method which integrates cosine similarity, norm clipping, historical trust memory and softmax-based weighting to mitigate the impact of conflicting or malicious client updates during federated aggregation. Evaluating proposed framework on the PathMNIST dataset with 30\% of persistent malicious clients both in IID and Dirichlet-based Non-IID federated environments to mimic heterogeneous medical learning environments. Comparing proposed aggregation method with FedAvg, Trimmed Mean and Coordinate Median.