Trust-gated median hybrid aggregation for byzantine-resilient federated learning on distributed EEG biomarkers for Alzheimer’s disease classification

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder whose early diagnosis remains challenging in distributed healthcare settings where direct sharing of patient data may be restricted. Federated learning enables collaborative model development while retaining raw data at local participants; however, its reliability can deteriorate under statistical heterogeneity and Byzantine clients that transmit malicious model updates. To address these challenges, this paper proposes a Trust-Gated Median Hybrid (TGMH) aggregation framework that integrates adaptive trust estimation, hard trust gating, coordinate-wise median protection, and adaptive hybrid fusion. The principal contribution lies in the server-side aggregation strategy for improving resilience to adversarial updates rather than in the development of a new diagnostic classifier. The framework is evaluated on 65 subjects from the OpenNeuro ds004504 EEG dataset using subject-independent five-fold outer cross-validation and simulated federated clients. TGMH is compared with FedAvg, FedProx, FedNova, FedAdam, SCAFFOLD, Coordinate Median, Trimmed Mean, Krum where theoretically feasible, and the Original Hybrid Trust–Trimmed framework. Robustness is evaluated under Sign Flip, Gaussian Poisoning, and Model Replacement attacks at 0, 20, and 40% Byzantine participation, together with an additional adaptive stealth stress test. Under the high-intensity 40% Model Replacement attack, TGMH substantially improves predictive performance relative to the Original Hybrid framework. Paired subject-level analysis further identifies six of 27 Byzantine-condition comparisons that remain significant after Holm correction, with large rank-biserial effects observed in several Model Replacement and Sign Flip comparisons. Subject-level bootstrap confidence intervals, component-wise ablation, trust-parameter sensitivity analysis, and scalability experiments provide complementary evidence regarding performance variability and the operating characteristics of the framework. Importantly, the adaptive stealth experiment shows that norm- and direction-constrained malicious updates can satisfy the current trust-gating criterion, indicating that the proposed defense does not provide universal robustness against adaptive adversaries. Scalability experiments with 5, 10, and 20 simulated clients further show that 50% partial participation reduces measured communication overhead by approximately 46–50% while producing only small within-configuration changes in mean Balanced Accuracy. Overall, the results demonstrate that TGMH provides a reproducible Byzantine-resilient aggregation framework with condition-dependent robustness advantages for federated EEG-based Alzheimer’s disease classification, while retaining raw subject data locally and without implying formal privacy guarantees.