Robust federated learning through decentralized adaptive oracle consensus

Federated learning (FL) is vulnerable to intermittent data poisoning and unreliable updates from untrusted participants. Existing blockchain-aided FL schemes provide immutable logging of model updates but still face three limitations: lack of temporal adversary modeling, reliance on costly on-chain computation, and communication overhead that restricts scalability. We propose State-Space Model-based Truth Discovery (SSMTD), a unified framework that addresses these challenges by integrating: (1) a Hidden Markov filter that assigns epoch-wise soft reputations to capture time-varying client reliability, (2) off-chain oracles that compute lightweight gradient-consistency metrics to avoid prohibitive gas costs, and (3) a decentralized oracle design that ensures robustness while maintaining sublinear communication growth as the number of clients increases. The inferred reputations are anchored on-chain. These reputations enable automated reward–penalty contracts that re-weight or exclude malicious participants in subsequent aggregation rounds. This combination of temporal reliability modeling, oracle-assisted off-chain computation, and scalable decentralized verification distinguishes SSMTD from prior approaches, which either neglect adversary dynamics or incur prohibitive overhead. Experiments on FMNIST, CIFAR-10, and LEAF under varying attack intensities and noise conditions show that SSMTD consistently outperforms reputation-based baselines, achieving 2.6–11.9 percentage-point improvements in F1-scores alongside simultaneous gains in precision and recall. System profiling further confirms stable throughput, sublinear latency growth, and minimal resource overhead, demonstrating strong scalability.