Differential Privacy with Selected Privacy Budget $\epsilon$ in a Cyber Physical System Using Machine Learning

Abstract

In contemporary data management practices, the adoption of Differential Privacy has emerged as a prevailing trend, offering an effective means to thwart an escalating array of query attacks. However, the implementation of Differential Privacy (DP) poses a nuanced challenge in determining the optimal privacy budget denoted by $\epsilon$. A small $\epsilon$ imparts formidable privacy fortification to the dataset, albeit rendering it scarcely utilizable and thus prone to abandonment due to severely compromised data utility. Conversely, an excessively large $\epsilon$ renders the dataset amenable for use, albeit at the cost of heightened susceptibility to privacy breaches via rudimentary attacks. Against this backdrop, the pivotal task becomes the judicious selection of an appropriate privacy budget value, one that harmonizes the imperatives of robust privacy protection and substantive data utility. This study endeavors to leverage the stochastic gradient descent (SGD) algorithm as a strategic approach to navigate this problem, aspiring to yield optimal resolutions to the presented challenge. A case study on real-world CPS testbed SWaT is conducted to demonstrate the feasibility of DP-enabled data privacy in time series data in a Historian server.