Reproduce “AutoFHE: Automated Adaption of CNNs for Efficient Evaluation over FHE” on OpenResearch

About this paper

Wei Ao, Vishnu Naresh Boddeti

AutoFHE automatically adapts Convolutional Neural Networks for secure and efficient inference over RNS-CKKS Fully Homomorphic Encryption, jointly optimizing layerwise mixed-degree polynomial approximations and the FHE evaluation architecture. This approach enables improved accuracy and significantly reduced inference latency compared to prior methods.

Key insights

• Automated, joint optimization of layerwise polynomial approximations and FHE architecture significantly outperforms manual methods, achieving superior accuracy-latency trade-offs by exploring a nuanced design space.
• Layer-specific sensitivities to approximation errors necessitate mixed-degree polynomials, as uniform approximation strategies are suboptimal, leading to highly non-uniform optimal multiplicative depth distributions across CNN layers.
• A hybrid approach that initializes from pre-trained ReLU network weights and fine- tunes with Polynomial-Aware Training (PAT) effectively combines the benefits of high accuracy from existing ReLU representations with the efficiency of tailored polynomial approximations.

Summary by alphaXiv.