A Privacy-Preserving Convolutional Neural Network Inference Framework for AIoT Applications

With the rapid development of the Artificial Intelligence of Things (AIoT), convolutional neural networks (CNNs) have demonstrated potential and remarkable performance in AIoT applications due to their excellent performance in various inference tasks. However, the users have concerns about privacy leakage for the use of AI and the performance and efficiency of computing on resource-constrained IoT edge devices. Therefore, this paper proposes an efficient privacy-preserving CNN framework (i.e., EPPA) based on the Fully Homomorphic Encryption (FHE) scheme for AIoT application scenarios. In the plaintext domain, we verify schemes with different activation structures to determine the actual activation functions applicable to the corresponding ciphertext domain. Within the encryption domain, we integrate batch normalization (BN) into the convolutional layers to simplify the computation process. For nonlinear activation functions, we use composite polynomials for approximate calculation. Regarding the noise accumulation caused by homomorphic multiplication operations, we realize the refreshment of ciphertext noise through minimal “decryption-encryption” interactions, instead of adopting bootstrapping operations. Additionally, in practical implementation, we convert three-dimensional convolution into two-dimensional convolution to reduce the amount of computation in the encryption domain. Finally, we conduct extensive experiments on four IoT datasets, different CNN architectures, and two platforms with different resource configurations to evaluate the performance of EPPA in detail.