Preserving Data Privacy during Neural Architecture Search for Edge AI Applications
Yi-Chuan Liang, Shan-Jung Hou, Shin-Wei Chiu, Shih‐Hao Hung · 2024
Neural Architecture Search (NAS) has become an important technique for finding efficient and accurate neural network models, especially for edge AI applications on power-constrained and specialized devices. NAS is typically done in the cloud, as edge devices do not possess the computing power to perform NAS, but revealing the actual dataset to the cloud causes privacy concerns. While homomorphic encryption (HE) can be used to preserve data privacy by uploading encrypted data, it results in high computational overhead for both the edge devices and cloud servers and reduced accuracy for nonlinear activation functions. This paper argues that functional encryption (FE) would be a good alternative to preserve data privacy during cloud-based NAS with significantly less overhead and better accuracy than HE. We present a system architecture for conducting NAS with FE and validate the argument with experiments and security analysis. On the server side, we use GPU to parallel the BSGS algorithm of the decryption operation After integrating FE into our framework and optimizing it, we utilized GPU parallel computation to accelerate decryption and improved performance by 3.5 times compared to the baseline design.