Towards the Design of Locally Differential Private Hardware System for Edge Computing
Kaito Taguchi, Kouichi Sakurai, Masahiro Iida · 2022
A challenging issue for edge computing is how to correct meaningful information from sensor data while keeping the privacy of the data and individuals. One approach is given by the recent work [Choi et al. “Guaranteeing Local Differential Privacy on Ultra-Low-Power Systems” 2018 ACM/IEEE 45th Annual International Symposium on Computer Architecture (ISCA)]. They point out that low resolution and fixed point characteristics of Ultra-Low-Power implementations may break privacy guarantees due to the low quality of noising. For overcoming this weakness, they introduce the techniques of resampling and thresholding. They also implemented in hardware to show the proposed method achieves both low overhead and high utility while keeping local differential privacy, with sensor/IoT benchmarks. Whereas, this research show some flaw in the existing work above and improve the method. We give the case in which infinite privacy loss still occurs because of low resolution even though the width of the output distribution is restricted, which was not investigated yet by the original paper of ISCA2018. Our major contribution is to propose an improvement to avoid privacy loss from low resolution and to enhance privacy protection in this circuit. Furthermore, we report the experimental results with software simulations, which guarantees the utility of our improving method. Finally, we discuss the power and limitation of our improved method with future challenging issues.