Staff Profile Pages

Evaluating the Robustness in Federated Learning for Medical Imaging

Authors: Soulb, H., Guelzim, I., Hirchoua, B., Nait-Charif, H.

Journal: 2025 6th International Conference on Data Analytics for Business and Industry Icdabi 2025

Publication Date: 01/01/2025

Pages: 265-270

DOI: 10.1109/ICDABI67967.2025.11547694

Abstract:

Federated Learning (FL) enables hospitals and research centers to collaboratively train AI models without sharingpatient data, offering strong potential for medical imaging. Yet, real-world adoption remains limited due to fragility in the faceof adversarial attacks, data heterogeneity, and noisy labels. We review robustness methods for medical imaging (2015-2025) and report two brain-MRI case studies comparing traditional baselines (FedAvg, FedProx) against a hybrid approach that augments FedAvg with BioGPT-based semantic weighting. We provide 95% confidence intervals, paired t/Wilcoxon tests, fairness metrics (max-min gap, stdev, Jain's index), and system costs (communication and LLM latency), with hardware, split statistics, and attack assumptions. Results show no significant gain of FedProx over FedAvg under short schedules; the FedAvg+LLM variant attains comparable accuracy and fairness with modest inference overhead. Under combined label-noise and Byzantine sign-flip stress, naïve averaging collapses toward chance, motivating robust aggregation plus lightweight semantic checks for trustworthy, scalable FL.

Source: Scopus