Addressing jointly privacy and ethical issues in responsible machine learning

点击翻译按钮获取中文摘要

The success of Machine Learning (ML) models is such that they are now ubiquitous in our society, which raises serious privacy and ethical issues (e.g., fairness and explainability), especially if their predictions are put into action in domains in which they can significantly affect individuals. There is a huge scientific literature that has studied in isolation, the concepts of privacy, fairness and interpretability in ML. However, very few works have explored the tensions, but also convergences, that emerge when addressing jointly these issues when designing and deploying ML models. The scope of my research program aims at answering this question by investigating when the achievement of these different objectives results in a positive sum game as well as the situations in which they conflict with each other, and a trade-off needs to be set. First, with respect to the intersection between privacy and fairness, the main challenge is to design fair models without collecting unnecessary sensitive attributes and while preventing privacy attacks on the training data or on the model. For example, I envision to adapt techniques coming from privacy such as anonymization mechanisms as well as data synthetization methods to also integrate fairness. Second, in the interactions between privacy and explainability, one of the main points of tension is that providing explanations necessarily reveal more information than simply outputting the prediction of the model. In this research direction, I propose to investigate which types of post hoc explanations among the existing ones offer the best robustness against privacy attacks. In addition of developing countermeasures to limit the success of such attacks, I will explore which families of interpretable models (e.g., linear models, decision trees or rule lists) offer the best protection against inference attacks. Third, at the crossings of fairness and explainability, I will work on the design of quantitative methods for evaluating the fairwashing risk in the context of posthoc explanations. I will also elaborate novel techniques for preventing fairwashing that aim at limiting the cheating possibilities of the adversary. To evaluate the practical efficiency of the methods developed, whenever possible they will be implemented in the form of prototypes, released in the form of an open-source library and experiments will be conducted with real datasets commonly used in the machine learning and fairness communities. The societal impact of the outcomes of my research program can be significant, as it has the potential to improve the privacy, fairness and explainability of models deployed in high-stake decision settings. In addition, the solutions developed will act as enablers by helping Canadian companies to implement models that are privacy-preserving and ethical by design. Finally, the research conducted will be done in cooperation with and contribute to the formation of HQP (i.e., PhD and master students).