Adversarial Robustness of Teacher-Student Networks

Teacher-student networks are typically used to distill knowledge from a large neural network to a smaller one. In this project, we investigated whether this paradigm can be used to create teacher-student ensembles having a certain level of adversarial robustness.

The results were published in a paper:

A. Bär, F. Hüger, P. Schlicht, and T. Fingscheidt. On the Robustness of Redundant Teacher-Student Frameworks for Semantic Segmentation (2nd place in the Best Paper Award rallye), in Proc. of CVPR - Workshops, Long Beach, CA, USA, Jun. 2019, pp. 1380 - 1388.

This project was a joint effort with the Volkswagen Group.