ESR Soumia Zohra El Mestari at 5th AAAI/ACM conference on Artificial Intelligence , Ethics and Society (AIES)
Early-Stage Researcher Soumia Zohra El Mestari attended the fifth AAAI/ACM conference on Artificial Intelligence, Ethics and Society (AIES) hosted at Keble college, university of Oxford in the United Kingdom from August 1st to 3rd 2022.
Soumia applied and was selected to attend the doctoral programme that was organized by the conference. As part of the student programme, Soumia participated in various workshops along with PhD students from different disciplines namely: Ethics, Law, Philosophy, and computer science.
The participants took part in many group working sessions along with mentors who are distinguished scientists in their fields. The workshops were followed by lunch roundtables with mentors.
During the last day of the conference, Soumia presented her PhD work around the usage of Privacy Enhancing Technologies(PETs) in machine learning pipelines as a tool to achieve the EU vision of trustworthy AI systems.
She first began by discussing the various conceptions of privacy in the context of data driven solutions, the different threat points throughout the machine learning pipeline and how these threat points can be exploited to maintain strong privacy attacks that can not only lead to heavy leakage bills but also may be invisible to detect. Then, the presentation also discussed the various PETs used to mitigate those attacks along with the pros and cons of each mitigation strategy.
Finally, Soumia concluded her presentation by stressing on the fact that the effective identification of the threat points can be a key element to solve some of the privacy issues. However, solving some privacy issues may come at the cost of other trustworthiness elements such as fairness and transparency. Hence, building a trustworthy AI pipeline is a challenging task since the acceptable tradeoff between those trustworthiness elements remains hard to measure.
Closing the debate, Tommaso highlighted that, contrary to what most data scientist would believe –that is, data protection and governance laws are obstacles to research—there are plenty of cases where laws such as the Open Data Directive or the Free flow of Non-Personal Regulation actually open up or give access to useful datasets – according to some specific conditions. Re- searchers, it seems, have more opportunities to access data than they think they have!
