Identity-based Secure and Efficient Intelligent Inference Framework for IoT-Cloud System

Abstract

The convolutional neural network (CNN) inference framework has been used in device-cloud systems to deploy near-end fast-response intelligent services. However, outsourcing data from devices to remote cloud for model training incurs security concerns, and existing inference models suffer from inefficiency and underperforming. In this paper, we design a novel framework for secure and efficient CNN inference based on IoT-edge-cloud collaboration. A two-layer identity-based cryptography scheme is designed to prevent sensor data and model parameters from leakage and tampering. A seed-filter-based model is leveraged to reduce model parameters for transmission and encryption, without sacrificing inference performance. The security analysis proves that our cryptographic algorithms can defeat Man-in-the-Middle attacks. Experimental results also indicate that the proposed framework can adapt to the efficiency requirements of edge computing without any compromise on the performance of machine learning tasks.