TRUST-WEIGHTED ADAPTIVE AGGREGATION WITH ANOMALY DETECTION FOR SECURE FEDERATED LEARNING AGAINST DATA POISONING

Federated Learning (FL) is a decentralised machine learning concept that enables collaborative model training by distributed clients without raw data centralisation for improved privacy. FL, though having these advantages, is very vulnerable to sophisticated data and model poisoning attacks where malicious clients poison updates to impair global performance. State-of-the-art protections such as static anomaly detection and Byzantine-robust aggregation are not robust against adaptive or covert attackers, particularly in heterogeneous (non-IID) data environments.This thesis offers a hybrid defence that combines unsupervised anomaly detection through autoencoder reconstruction loss with dynamic trust-weighted aggregation. The system computes the trust scores of the clients from live monitoring of deviations from learned benign update patterns and uses the scores to dynamically weight client contribution on a per-aggregation-round basis. The technique supports real-time detection and incremental defence against adversarial behaviour with no prior attack method knowledge and no client rejection at all.