Sergio A. C. Romero, Renato Candido, Magno T. M. Silva. Local Graph-Homomorphic Privacy in Distributed Neural Networks In: 13th SIMBig International Conference on Information Management and Big Data, 2026, Arequipa. Proceedings of SIMBig’2026, 2026.

Abstract

The growing concern over the privacy of sensitive data in the training of machine-learning models has driven the development of robust protection techniques. This work investigates privacy-preserving methods for distributed neural networks, with emphasis on adaptive diffusion techniques. Random noise, global homomorphic, and local homomorphic perturbations are compared in terms of classification utility and empirical information leakage. Experimental results on two real-world datasets show that local homomorphic noise preserves classification accuracy close to that of the non-private model. Moreover, the improved Deep Leakage from Gradients (iDLG) attack, local homomorphic noise substantially degrades the quality of the reconstructed data, as indicated by an increase in reconstruction error and a reduction in the structural similarity index. These results indicate a favorable empirical privacy-utility trade-off for local graph-homomorphic perturbation.

Keywords

Machine learning, Privacy, Distributed training, Sensitive data.

Downloads

• PDF