| Speaker : | Patrick Loiseau |
| EURECOM | |
| Date: | 11/03/2015 |
| Time: | 2:00 pm - 3:00 pm |
| Location: | LINCS Meeting Room 40 |
Abstract
Statistical learning methods developed in the last decades have proved very useful for many applications. However, most algorithms were developed under the assumption that the data is independent from the algorithm. This is no longer true in applications where data is generated or provided by (human) strategic agents. As more and more modern applications indeed learn from data provided by external parties, it becomes increasingly crucial to account for the data-provider’s incentives in order to design well-performing learning algorithms in practice.In this talk, we use game theory to model explicitly the economic incentives of the learner and of the agents generating data. We show how, by analyzing the game, it is possible to design learning algorithms that work better because they account for the agents’ incentives. Specifically, we illustrate this in two examples with applications in security and privacy: (i) Classification of malicious behavior: In this example, a defender runs a classification algorithm to detect malicious behavior while an attacker chooses his attack strategy to balance the utility of the attack and the probability of being detected. We obtain analytic results that show how to optimally perform the classification given the attacker’s utility. (ii) Linear regression from agents-provided data: In this example, we study the case of an analyst who tries to infer a model through linear regression from data provided by individuals who add noise to protect their privacy (this technique is called “local privacy”). We again derive results on optimal linear regression and on the price of anarchy of the system that provide methods to better learn from privacy-conscious users by incentivizing them to provide higher-quality data.