Dynamic Sparsification in Secure Gradient Aggregation for Federated Learning

Abstract

Secure aggregation is a critical component of privacy-preserving federated learning. However, existing fixed-sparsity approaches often incur unnecessary communication overhead. We present DynamicSecAgg, a novel framework that introduces dynamic sparsity while preserving coordinate-level privacy. Our method achieves significant improvements in communication efficiency while maintaining — and in some cases improving — model accuracy across both IID and non-IID user distributions. The framework maintains information-theoretic privacy guarantees via adaptive gradient thresholding and polynomial-based aggregation, proving particularly effective under heterogeneous data settings. These results establish dynamic sparsity as a key optimization for efficient and privacy-preserving federated learning.