Training Large Scale Polynomial CNNs for E2E Inference over Homomorphic Encryption
This work addresses the adoption bottleneck for HE-based solutions in privacy-preserving machine learning by enabling efficient encrypted inference for large models, though it is incremental in adapting existing architectures to polynomial constraints.
The paper tackles the challenge of training large-scale polynomial CNNs for inference under Homomorphic Encryption (HE), which is limited by polynomial-only operations, and achieves promising accuracy on encrypted ImageNet samples with models like ResNet-152 and ConvNeXt. It also adapts CLIP for secure zero-shot prediction, demonstrating robustness at the intersection of HE and transfer learning.
Training large-scale CNNs that during inference can be run under Homomorphic Encryption (HE) is challenging due to the need to use only polynomial operations. This limits HE-based solutions adoption. We address this challenge and pioneer in providing a novel training method for large polynomial CNNs such as ResNet-152 and ConvNeXt models, and achieve promising accuracy on encrypted samples on large-scale dataset such as ImageNet. Additionally, we provide optimization insights regarding activation functions and skip-connection latency impacts, enhancing HE-based evaluation efficiency. Finally, to demonstrate the robustness of our method, we provide a polynomial adaptation of the CLIP model for secure zero-shot prediction, unlocking unprecedented capabilities at the intersection of HE and transfer learning.