DMTT federated learning protocol resists 80% adversarial attacks with 0.862 accuracy

Researchers Shubham Vaishnav, Murtaza Rangwala, Ali Beikmohammadi, Sindri Magnússon, and Rajkumar Buyya have introduced DMTT (Dynamic MURMURA with Trusted Topology), a decentralized federated learning (DFL) protocol designed to counter adversaries who poison both model updates and the topology information devices use to select collaborators. Built on the MURMURA algorithm, DMTT extends evidential deep learning to dynamically weight distribution-mismatched peers while also defending against topology-manipulation attacks on time-varying graphs. Each device maintains a confidence-weighted local topology view using link-reliability estimates, signed topology claims, witness corroboration, and a Beta-distributed source-trust model. Aggregation occurs only over a trust-screened collaborator set, using a composite score that fuses model compatibility, topology trust, and link reliability.

DMTT's theoretical contribution is a proof that screened mixing matrices confine Byzantine influence to a bounded residual δ_max, which vanishes under perfect screening. Empirically, the protocol was tested on UCI HAR and PAMAP2 datasets, each partitioned across 100 mobile clients with Dirichlet heterogeneity. DMTT maintained honest-node accuracy above 0.862 (UCI HAR) and 0.829 (PAMAP2) across adversary fractions ranging from 10% to 80%—nearly matching no-attack accuracy at low fractions and gracefully degrading toward local-only performance at extremes. In contrast, static and dynamic FedAvg collapsed to chance at every fraction, and robust aggregators like Krum, BALANCE, and UBAR failed to consistently beat a local-only baseline. The implementation runs coordinator-free via a ZeroMQ backend on the Melbourne Research Cloud, with surviving Byzantine aggregation weight empirically zero, confirming δ_max=0.