Machine learning models are being extensively trained on sensitive human data (such as pictures, videos and patient health records) and are being publicly deployed as a service. With such systems in place, privacy of the individuals involved in the model training becomes a real concern. While differential privacy provides a solution to this problem, there is always a privacy-utility trade-off when training models privately, which is not well understood. Often the practical deployments choose the model utility over privacy, that may lead to indiscernable privacy vulnerabilities. One such example of privacy vulnerability is whether an adversary can identify a particular individual in the training data. Also, what type of individuals are more vulnerable to such attacks? This question is also directly related to the problem of fairness. In light of such vulnerabilities, what privacy parameters should the ML deployments use to mitigate the risks? Our project tries to shed light on these questions.
Analysis of Private ML Models
Membership inference attacks are effective even for skewed priors.
What seems safe, might not be safe in practice.
We propose novel membership inference attack and a threshold selection procedure to improve the existing attacks.
We compare the privacy leakage of ML models trained with different differential privacy relaxations and different privacy budgets.