PAPER: Privacy-Preserving ResNet Models using Low-Degree Polynomial Approximations and Structural Optimizations on Leveled FHE

Recent work has made non-interactive privacy-preserving inference more practical by running deep Convolution Neural Network (CNN) with Fully Homomorphic Encryption (FHE). However, these methods remain limited by their reliance on bootstrapping, a costly FHE operation applied across multiple layers, severely slowing inference. They also depend on high-degree polynomial approximations of non-linear activations, which increase multiplicative depth and reduce accuracy by 2-5% compared to plaintext ReLU models. In this work, we focus on ResNets, a widely adopted benchmark architecture in privacy-preserving inference, and close the accuracy gap between their FHE-based non-interactive models and plaintext counterparts, while also achieving faster inference than existing methods. We use a quadratic polynomial approximation of ReLU, which achieves the theoretical minimum multiplicative depth for non-linear activations, along with a penalty-based training strategy. We further introduce structural optimizations such as node fusing, weight redistribution, and tower reuse. These optimizations reduce the required FHE levels in CNNs by nearly a factor of five compared to prior work, allowing us to run ResNet models under leveled FHE without bootstrapping. To further accelerate inference and recover accuracy typically lost with polynomial approximations, we introduce parameter clustering along with a joint strategy of data encoding layout and ensemble techniques. Experiments with ResNet-18, ResNet-20, and ResNet-32 on CIFAR-10 and CIFAR-100 show that our approach achieves up to 4x faster private inference than prior work with comparable accuracy to plaintext ReLU models.