Cryptographic Sparsification for Privacy-Preserving Collaborative Learning in Zero-Trust Edge Architectures

Article Information

Abstract

Although decentralized machine learning paradigms safeguard private user data by exchanging model updates rather than raw records, advanced gradient inversion techniques have shown that sensitive source inputs can still be reconstructed with high fidelity. Standard cryptographic countermeasures, such as Homomorphic Encryption (HE), prevent untrusted server nodes from accessing gradients directly, yet they introduce severe compute and transmission bottlenecks—frequently inflating payload dimensions by 10x to 50x. To mitigate this constraint, we present CryptoSparsity, a novel framework tailored for zero-trust edge environments. CryptoSparsity combines dynamic threshold-based gradient pruning with additive homomorphic operations. By evaluating the relative magnitude of model updates during each training step, the architecture eliminates redundant parameters (filtering up to 95% of non-essential weight variations) prior to encrypting the remaining residual vectors using an efficient Paillier cryptosystem. Evaluated across 150 physical edge nodes (comprising NVIDIA Jetson devices and Raspberry Pi 4 systems) training Deep Neural Networks, CryptoSparsity decreased total encrypted packet size by 88.4% and cut client-side encryption processing time by 82.1%, while maintaining robust defense against gradient-based data recovery attacks.

Keywords

DOI Active Record

This scholarly article has been successfully registered and assigned a persistent DOI through DOIGLOBAL.

Research Tools

Share This Research

Recommended Citation