TrustLite-FL: multi-statistic trust aggregation for robust federated learning in consumer IoT

Federated learning (FL) trains shared models on consumer IoT devices without moving raw data off the device, but poisoning can compromise it. We propose TrustLite-FL , a trust-weighted aggregation framework that mitigates several poisoning behaviors. It scores client updates by entropy, magnitude and direction relative to the round, smooths those scores over time, clips update norms, and logs suspicious behavior for audit. Its aggregation and scoring scale linearly with clients and model parameters. We prove that update entropy is invariant to positive scaling and sign flips, so entropy alone cannot detect them. We evaluate it in simulation on MNIST, Fashion-MNIST and CIFAR-10, benchmark proxies for consumer device data, with non-IID clients, five attacks plus a clean setting, 30% adversaries, seven baselines and three seeds. It has the highest mean accuracy, or matches it, in 9 of 15 attacked settings, and is within one percentage point of the best in 13. Ablation shows that the norm clip provides most of this robustness. The trust layer excludes every Gaussian noise adversary, excluding at most 2.9% of benign clients, but no sign flipping or ALIE adversary. Aggregation is 11 times slower than FedAvg at 1,000 clients but faster than Krum from 500 clients.