Human Aligned Compression for Robust Models

Samuel Räber, Andreas Plesner, Till Aczel +1

Adversarial attacks on image models threaten system robustness by introducing imperceptible perturbations that cause incorrect predictions. We investigate human-aligned learned los…