Adaptive attacks on machine learning models

What is it about?

In this work we propose an optimized strategy to attack machine learning models for classification, e.g., to make malware look like benign software to the classifier. The strategy finds the optimal balance between two competing actions available to the attacker, thus boosting the effectiveness of the final attack.

Featured Image

Photo by Michael Geiger on Unsplash

Photo by Michael Geiger on Unsplash

Why is it important?

Our work proposes a more realistic attack strategy than the state of the art, where the attacker is forced to follow a fixed attack pattern not representative of real-world scenarios. We experimentally show that the increase in the attack effectiveness enabled by our approach is significant.