Game Theory and Machine Learning
Despite the success of machine learning, many modern applications now involve data generated or provided by selfinterested strategic agents whose objectives depend on the outcome of the learning algorithms. In such cases, standard learning algorithms tend to perform poorly. This is the case for instance in security, where data is generated by attackers who try to evade detection or in applications that involve personal data revealed by privacyconscious users who try to limit their privacy risks when revealing data. At EURECOM, we work on designing learning algorithms that work better in those cases by using game theory to model the interactions between learning agents and agents generating or providing data and to design algorithms that work optimally given the agents' incentives. Topics include:
 Game theory for adversarial classification in security
 Learning in defense resource allocation games (Blotto)
 Learning from personal data of privacyconscious users
 Statistical learning theory with strategic data
Selected publications:

L. Dritsoula, P. Loiseau, and J. Musacchio. A gametheoretic analysis of adversarial classification. arXiv:1610.04972, 2016. [ bib  pdf  http ]

V. Kamble, P. Loiseau, and J. Walrand. Regretoptimal strategies for playing repeated games with discounted losses. arXiv:1603.04981, 2016. [ bib  pdf  http ]

M. Chessa, J. Grossklags, and P. Loiseau. A gametheoretic study on nonmonetary incentives in data analytics projects with privacy implications. In Proceedings of the 28th IEEE Computer Security Foundations Symposium (CSF), 2015. [ bib  pdf  http ]

S. Ioannidis and P. Loiseau. Linear regression as a noncooperative game. In Proceedings of the 9th conference on Web and Internet Economics (WINE), 2013. [ bib  pdf  http ]

L. Dritsoula, P. Loiseau, and J. Musacchio. Computing the nash equilibria of intruder classification games. In Proceedings of the third Conference on Decision and Game Theory for Security (GameSec), 2012. [ bib  http ]