Attacking Machine Learning Models for Social Good

Vibha Belavadi, Yan Zhou, Murat Kantarcioglu, Bhavani Thuriasingham

Abstract: As machine learning (ML) techniques are becoming widely used, awareness of the harmful effect of automation is growing. Especially, in problem domains where critical decisions are made, machine learning-based applications may raise ethical issues with respect to fairness and privacy. Existing research on fairness and privacy in the ML community mainly focuses on providing remedies during the ML model training phase. Unfortunately, such remedies may not be voluntarily adopted by the industry that is concerned about the profits. In this paper, we propose to apply, from the user’s end, a fair and legitimate technique to “game” the ML system to ameliorate its social accountability issues. We show that although adversarial attacks can be exploited to tamper with ML systems, they can also be used for social good. We demonstrate the effectiveness of our proposed technique on real world image and credit data.