Byzantine robust federated learning for heterogeneous brain MRI using multisignal gradient fingerprinting and...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Imaging

Background:

• Federated learning (FL) enables collaborative model training across institutions without data centralization.
• FL is vulnerable to malicious clients and non-independent and identically distributed (non-IID) data heterogeneity.
• Securing FL in sensitive domains like medical imaging is crucial for reliable collaborative analysis.

Purpose of the Study:

• To develop a trust-aware federated learning framework for brain MRI analysis.
• To enhance Byzantine robustness against malicious clients and data heterogeneity.
• To improve the security and reliability of collaborative medical image analysis.

Main Methods:

• Proposed a trust-aware FL framework combining multi-signal gradient fingerprinting and adaptive aggregation.
• Utilized a six-dimensional fingerprint including VAE reconstruction error, cosine similarity, peer similarity, gradient norm, sign consistency, and Monte Carlo Shapley contribution.
• Integrated a dual-attention module and reinforcement learning controller with FedBN-P for robust aggregation.

Main Results:

• Achieved high detection F1 scores (above 0.98) for gradient-scaling attacks.
• Demonstrated robustness against up to 40% malicious clients with minimal model accuracy degradation.
• Framework showed a +8.8% wall-clock overhead compared to FedAvg with identical communication volume.
• Ablation studies confirmed the effectiveness of individual defense components (VAE fingerprinting, Shapley values, RL controller).

Conclusions:

• The proposed trust-aware FL framework effectively enhances Byzantine robustness in brain MRI analysis.
• The defense-in-depth design provides layered security against various adversarial attacks.
• Further validation on larger, multi-site federated cohorts is necessary before clinical deployment.