A practical homomorphic encryption approach for GDPR-compliant machine learning full training protocol

This work introduces a novel privacy-preserving full training protocol for deep learning models using external data. It addresses the critical, yet under-explored, challenge of data privacy, especially under regulations like GDPR. While homomorphic encryption has been suggested as a breakthrough for utilizing private external datasets in machine learning, naively applying it to model training from scratch is impractical due to huge computational costs. The proposed method strategically encrypts only the minimal layers required to prevent privacy breaches. This approach is carefully designed to minimize the computational overhead of homomorphic encryption, making it efficient and scalable for larger models. Comprehensive analysis on benchmark datasets (ResNet-20/110 with CIFAR-10/100) confirms compliance with GDPR requirements without significant loss of accuracy. Experiments show a 1000x reduction in training time compared to models encrypted across all layers, which is the first published benchmark as far as we know.