Security, Privacy, and Transparency Guarantees for Machine Learning Systems

Machine learning (ML) is transforming a wide range of applications, promising to bring immense economic and social benefits. However, it also raises substantial security, privacy, and transparency challenges. ML workloads indeed push companies toward aggressive data collection and loose data access policies, placing troves of sensitive user information at risk if the company is hacked. ML also introduces new attack vectors, such as adversarial example attacks, which can completely nullify models’ accuracy under attack. Finally, ML models make complex data-driven decisions, which are opaque to the end-users, and difficult to inspect for programmers. In this dissertation we describe three systems we developed. Each system addresses a dimension of the previous challenges, by combining new practical systems techniques with rigorous theory to achieve a guaranteed level of protection, and make systems easier to understand. First we present Sage, a differentially private ML platform that enforces a meaningful protection semantic for the troves of personal information amassed by today’s companies. Second we describe PixelDP, a defense against adversarial examples that leverages differential privacy theory to provide a guaranteed level of accuracy under attack. Third we introduce Sunlight, a tool to enhance the transparency of opaque targeting services, using rigorous causal inference theory to explain targeting decisions to end-users.