Efficient FHE-Based Privacy-Enhanced Neural Network for Trustworthy AI-as-a-Service

AI-as-a-Service has emerged as an important trend for supporting the growth of the digital economy. Digital service providers make use of their vast amount of customer data to train AI models (such as image recognition, financial modelling and pandemic modelling etc) and offer them as a service on the cloud. While there are convincing advantages for using such third-party models, the fact that model users are required to upload their data to the cloud is bound to raise serious privacy concerns, especially in the face of increasingly stringent privacy regulations and legislation. To promote the adoption of AI-as-a-Service while addressing privacy issues, we propose a practical approach for constructing privacy-enhanced neural networks by designing an efficient implementation of fully homomorphic encryption. With this approach, an existing neural network can be converted to process FHE-encrypted data and produce encrypted output which are only accessible by the model users, and more importantly, within an operationally acceptable time (e.g., within 1 s for facial recognition in typical border control systems). Experimental results show that in many practical tasks such as facial recognition, text classification and so on, we obtained the state-of-the-art inference accuracy in less than one second on a 16 cores CPU.