CryptDNN: A Fast Privacy-Preserving Deep Neural Network Inference Architecture Based on Cloud-Edge-Client Collaboration
Dong Li, Anupam Chattopadhyay, Qianyu Li, Qingguo Lü, Jiahui Wu, Tao Xiang, Xiaofeng Liao · IEEE Transactions on Network Science and Engineering · 2025
Deep neural networks (DNNs) with fully homomorphic encryption (FHE) are gaining traction in Internet of Things (IoT) applications requiring robust protection of data inference and model privacy. However, the existing implementations typically replace activation functions with approximate polynomial functions, which compromises the accuracy and efficiency of machine learning. To overcome these limitations, we introduce CryptDNN, a fast privacy-preserving DNN inference architecture based on cloud-edge-client collaboration. In CryptDNN, the DNN model and the data in IoT are deployed in the cloud server, the client, and the edge server in a distributed manner. Specifically, we first encrypt the client's data by using FHE, then divide the DNN into two components: a linear module encrypted with FHE which is stored in the cloud server, and a non-linear module represented as an encrypted look-up table via searchable symmetric encryption (SSE), which is stored in the edge server, thus avoid the need for low-degree polynomial approximations of the activation function under ciphertext environments. The security analysis confirms that the CryptDNN robustly safeguards the confidentiality of both the data and the model. Experimental results on the realistic MNIST and CIFAR-10 datasets also demonstrate that the CryptDNN achieves high inference accuracy and enhanced computational efficiency.