SeP2CNN: A Simple and Efficient Privacy-Preserving CNN for AIoT Applications
Xiao Tao, Zhenyong Zhang, Kuan Shao, Hao Li · 2024
As convolutional neural networks (CNNs) have shown excellent performance in various inference tasks, it has become increasingly critical to enable Artificial Intelligence of Things ($A$IoT) systems to run CNN-based applications. However, using CNN-based applications in many AIoT scenarios raises privacy concerns. Protecting the inference phase can preserve the user's privacy, which is an urgent issue according to worldwide rules and policies. In this paper, we propose SeP2CNN, a simple and efficient fully homomorphic encryption (FHE) method to realize the privacy-preserving CNN inference in pervasive AIoT applications. By carefully combining cryptography and deep learning, we compute ReLU with a composite polynomial as the activation function over ciphertexts. To improve the efficiency, we transform the three-dimensional convolution into a two-dimensional convolution to simplify the calculation. Finally, through extensive experiments on multiple datasets, we validate and evaluate the performance of SeP2CNN.