Federated & Private ML

AR-FL: Anomaly-Resilient Federated Learning

AR-FL: Anomaly-Resilient Federated Learning for Robust Distributed Systems

Federated learning that notices when a client starts sending strange updates and quietly turns its volume down.

The problem

Client updates can be corrupted by faulty data, unreliable devices or adversaries, and fixed anomaly thresholds do not adapt to how noisy a round is.

The approach

A Weighted Adaptive Isolation Forest flags anomalous updates, with its contamination rate tuned from validation-loss degradation, followed by softmax-weighted aggregation.

The outcome

Lower error than existing aggregation methods on real-world time-series prediction, especially at high anomaly rates.

In detail

Developed a federated learning framework for distributed environments where client updates may be corrupted by faulty data, unreliable devices or adversarial behaviour. AR-FL combines a Weighted Adaptive Isolation Forest with validation-driven contamination adjustment and softmax-based aggregation, allowing the global model to reduce the influence of anomalous updates as conditions change.

Built with

• Federated Learning
• Anomaly Detection
• Robust Aggregation

More in Federated & Private ML

• IRT-FedLoRA: Reliability-Aware Federated Fine-Tuning
• ClusterGuardFL: Secure Cluster-Weighted Client Aggregation
• DP-FedMeta: Differentially Private Federated Meta-Learning

Next project: DP-FedMeta: Differentially Private Federated Meta-Learning

By Kanishka Ranaweera, PhD, Machine Learning Engineer in Melbourne, Victoria, Australia. hello@kanishkabuilt.me