Perceptual Encryption-based Privacy-Preserving Deep Learning in Internet of Things Applications
Published in IEEE International Conference on Information and Communication Technology Convergence (ICTC), 2022
Abstract
In the Internet of Things (IoT) ecosystem, cloud computations are widely utilized to train a deep learning model on public datasets. The trained model can then be deployed on edge devices as an inference engine to facilitate various IoT applications in real-time. For consistent accurate predictions, the model needs to be retrained on the most recent data. This necessitates the data share between cloud servers and edge devices. However, the data acquired by IoT end devices usually consists of sensitive information and sharing them with cloud services provider results in users' privacy issues. In addition, the exchange of large volume of data requires high bandwidth. This study proposes an extension of block-based perceptual encryption (PE) algorithm to enable DL computations in encryption domain in order to protect users’ privacy. For this purpose, four different extensions of baseline PE methods were analyzed in terms of compression savings, encryption efficiency and privacy-preserving DL performance. The analyses have shown that when global contents are preserved in each color channel, the extended PE method offers better security and preserves compression savings with only 3% drop in accuracy as compared to baseline method.