On Sep 21, 2021, we had the pleasure to visit our partner EDF in Paris Palaiseau! Here is another highlight presentation:
Hassan CHAITOU, Security risk optimization for learning on heterogeneous quality data
A classifier is a component used in the automation of “decision-making” or complex data abstraction: intruder detection, speed limitation extraction. For an efficient classifier, the training must be on a large volume of data and be renewed over time by integrating or revoking certain learning data. From a security point of view, this process represents a risk since it offers the attacker various ways of degrading classifier performance (either by forcing classifications mischievous, either by randomly degrading its performance). These two types of attacks require more or less effort from the attacker.
This risk is exacerbated when data comes from sources (network equipment, organizations) corresponding to heterogeneous trust levels. Hassan’s thesis aims at controlling the risk associated with this update via game theory in the case where the confidence in the learning data is not homogeneous.