Group verifiable secure aggregate federated learning based on secret sharing.

Area of Science:

• Computer Science
• Machine Learning
• Cybersecurity

Background:

• Federated learning (FL) addresses data silos and raw data security but faces privacy leakage and server tampering risks.
• Existing privacy-preserving methods in FL often incur high computational and communication costs, limiting their use in resource-constrained environments.

Purpose of the Study:

• To propose an efficient and secure aggregation scheme, GVSA (Grouped Vector Secret Aggregation), for federated learning.
• To enhance privacy, resilience to user dropouts, and efficiency in federated learning systems.

Main Methods:

• GVSA employs a masking technique for local model privacy and secret sharing for resilience against user dropouts.
• The scheme utilizes a dual aggregation approach with lightweight validation tags for accuracy verification.
• A grouping strategy is implemented to reduce computational load on users and the server.

Main Results:

• GVSA demonstrates high security and effective model accuracy preservation.
• Compared to FedAvg, GVSA shows only ~7% additional computational overhead.
• GVSA achieves ~2.3x faster training speed than other secure aggregation schemes at the same security level.

Conclusions:

• GVSA offers an efficient and secure solution for federated learning, particularly suitable for resource-limited environments.
• The proposed scheme balances strong privacy guarantees with practical performance improvements.