Secure Aggregation for Privacy-Preserving Federated Learning on Clinical EEG Data

This study addresses the privacy risks inherent in federated learning (FL) for clinical EEG data, where model updates can potentially leak sensitive patient information. The authors propose a secure aggregation (SecAgg) framework designed for cross-silo healthcare environments. The system allows institutional clients to train models locally and contribute to a global model without the central server ever seeing individual, unmasked updates. The framework is built on the Flower federated learning platform and incorporates several cryptographic techniques, including threshold secret sharing, graph-based communication, and local update clipping.

The authors implement and evaluate four distinct protocol variants to balance security and performance:

• Semi-Honest (Π1): Provides the baseline protection against an honest-but-curious server with minimal computational overhead.
• Malicious (Π2): Adds Merkle tree commitments, proof verification, and consistency checks to prevent the server from presenting inconsistent views or forging participation claims.
• Semi-Honest with Auxiliary-Notary (Π3): Extends the semi-honest variant with a lightweight verification mechanism that allows clients to confirm the integrity of the final aggregate.
• Malicious with Auxiliary-Notary (Π4): Combines the robust malicious-setting safeguards with the auxiliary-notary verification for maximum integrity and consistency.

The framework was evaluated using EEG data from the TUH EEG Corpus across 10, 40, and 70-client configurations. The results demonstrate that while secure aggregation is feasible for clinical EEG tasks, there is a clear trade-off between security and efficiency. The semi-honest variant (Π1) imposes the lowest computational and communication burden. In contrast, the malicious and auxiliary-notary variants introduce significant overhead due to the additional cryptographic proofs and consistency-checking steps required to ensure protocol integrity. Despite these costs, the framework remains compatible with standard federated training pipelines.