Privacy-Preserving Detection of Helmet and Mask Wearing with Fully Homomorphic Encryption: Towards a Secure Inference Approach

Jia-Lin Chan, Wun-She Yap, Denis Chee-Keong Wong, Bok-Min Goi, Wai-Kong Lee

Abstract: In the modern era, technological advancements serve not only to optimize operational efficiency but also to fortify security measures across various domains. For instance, deep learning (DL) is utilized for detecting suspicious activities, such as wearing a mask and/or helmet within the ATM premises, to reinforce public safety and banking security. These trained AI models are commonly deployed on cloud servers, enabling users to process their data using pretrained models, easing the burden of computational overhead. However, privacy concerns remain prevalent as ATM users may not want to disclose their sensitive data to third parties. Fully Homomorphic Encryption (FHE) offers a solution by enabling data computation within the encrypted domain, thereby ensuring user privacy. With FHE, users can encrypt their data before transmitting it to the cloud for inferences, thus relieving their privacy concern, because the cloud learns nothing from the received encrypted data. The inference results are sent back to the users for decryption and consumption. This paper presented a privacy-preserving DL system for inference, which utilizes the Cheon-Kim-Kim-Song (CKKS) scheme to protect LeNeT-5, which is a DL model trained to detect helmet-mask wearing. The research findings demonstrate the promising potential of FHE inference for privacy-preserving applications, with a 0.48% overall accuracy loss compared to plaintext inference. Notably, the system achieves a commendable accuracy rate of 91.17% for helmet detection and 92.31% for mask detection in the encrypted domain. The proposed solution can be used to protect the privacy of bank users, at the same time provide the surveillance service to the bank industry.

External IDs:dblp:conf/rtsi/ChanYWGL24