Investigating Membership Inference Attacks under Data Dependencies. Thomas Humphries, Simon Oya, Lindsey Tulloch, Matthew Rafuse, Ian Goldberg, Urs Hengartner, and Florian Kerschbaum. In Computer Security Foundations Symposium (CSF), 2023.
Machine Learning Privacy
Machine learning technologies are everywhere. Here’s the privacy angle of all this: when one builds a machine learning model using a privacy-sensitive training dataset, the model will memorize some privacy-sensitive information. An attacker with access to that model (even if it is just black-box query access) can learn a lot of information about the training data. The most basic attack in this setting is called membership inference attack, where an attacker gets access to a model and a data sample, and wants to guess whether that sample was in the model’s training set. More complex attacks include model inversion, where the attacker reconstructs training samples just by querying the model! Sadly, due to the complexity of machine learning (it’s called machine learning, after all), preventing these attacks is really hard!
I am interested in understanding the privacy implications of training machine learning models on sensitive data, developing tools to quantify their privacy leakage, and building defenses to limit such leakage. As with the rest of my research, I try to avoid purely empirical or heuristic results, and instead aim get some solid understanding of why things work the way they do. This is sadly somewhat hard in machine learning, but we have to try!
The only paper I have (so far) in the topic should give you a good idea of the type of questions I like to study in this area: