Privacy Attacks, breaking the anonymization to refine the privacy evaluation

In recent years, the volume of data generated has followed an exponential trend, more and more information are collected and stored and companies can process these data to extrapolate useful information for their business. The importance of privacy is growing in people's minds as they became aware of the potential adversarial use cases that unprotected data could lead to, and countries followed this trend by rapidly adapting to Data Regulations. To comply with these regulations, companies must rely not only on adequate data processing techniques but also on how to detect these risks in the first place. This project, carried out within the Accenture Lab AI division, is born to explore the state of the art in terms of privacy attacks and corresponding defence mechanisms to expand the Automated Privacy & Value Assessment Tool (APAT) already developed by the department. The project aims to identify the exploit that a dataset is subject to after being published, and what an adversary can infer from this release with little or no background knowledge. The first phase of exploration with a special focus on Membership Inference Attack is followed by an implementation of these risk evaluations inside the existing tool to improve its capabilities of detecting privacy leaks. The work is concluded with the development of a web application that showcases how a client can test its anonymized dataset against membership inference.