Federated & Private ML

IRT-FedLoRA: Reliability-Aware Federated Fine-Tuning

Item Response Theory-Guided Aggregation for Robust Federated Low-Rank Adaptation

A way to fine-tune large models across many organisations without trusting every participant equally. Each client earns a reliability score over time, and the model listens more to the clients that have proven consistent.

The problem

Federated LoRA fine-tuning usually averages every client’s update the same way, so noisy, inconsistent or adversarial clients drag the shared model down.

The approach

Item Response Theory treats each client as a test-taker and each training round as a question. Reliability is a latent trait inferred from update alignment, validation gains and gradient consistency, then used to weight each client’s LoRA update.

The outcome

A principled, interpretable and lightweight aggregation rule that emphasises trustworthy clients and suppresses unreliable ones.

In detail

Proposed IRT-FedLoRA for robust federated fine-tuning of large models using Low-Rank Adaptation. Each client is modelled as a respondent and each training round as an item, allowing latent client reliability to evolve over time. Reliability is inferred from update alignment, validation improvement and gradient consistency, then used to adaptively weight contributions during aggregation.

Built with

• LLMs
• Federated Fine-Tuning
• Item Response Theory

More in Federated & Private ML

• ClusterGuardFL: Secure Cluster-Weighted Client Aggregation
• AR-FL: Anomaly-Resilient Federated Learning
• DP-FedMeta: Differentially Private Federated Meta-Learning

Next project: Pneumonia Detection with Vision Transformers

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