🤖 AI Summary
This work addresses the challenge of privacy-preserving training of multi-layer neural networks under pure homomorphic encryption (HE), where no trusted third party or hybrid security protocols are assumed. Methodologically, it proposes the first end-to-end trainable three-layer fully encrypted neural network framework—supporting both regression and classification—built upon the CKKS HE scheme. Key technical contributions include: efficient ciphertext matrix operations; Chebyshev polynomial approximations for non-linear activation functions; and a secure gradient clipping mechanism with adaptive rescaling to ensure training stability. Experiments on MNIST (classification) and UCI regression datasets demonstrate that model accuracy (classification) and mean squared error (regression) closely match those of plaintext training. To the best of our knowledge, this is the first work to realize complete, end-to-end training of a three-layer neural network entirely within the pure HE paradigm, thereby substantially advancing the practical applicability of HE in deep learning.
📝 Abstract
In this manuscript, we consider the problem of privacy-preserving training of neural networks in the mere homomorphic encryption setting. We combine several exsiting techniques available, extend some of them, and finally enable the training of 3-layer neural networks for both the regression and classification problems using mere homomorphic encryption technique.