Our paper “Shedding More Light on Robust Classifiers Under the Lens of Energy-Based Models” at ECCV'24!

Shed Light on Robust Classifiers

This work reinterprets a robust discriminative classifier as Energy-based Model (EBM) and offer a new take on the dynamics of adversarial training. Our research introduces novel theoretical and practical insights demonstrating how analyzing the energy dynamics of adversarial training (AT) enhances our understanding. We also present a instance reweighting scheme, that weights samples based on their energy while adversarial training, thereby improving the model’s robustness. This project also explores the generative capabilities of robust classifiers under the lens of energy-based models (EBMs). We demonstrate that robust classifiers exhibit varying intensities and qualities in their generative capabilities. Furthermore, we propose a straightforward method to enhance this capability.

More info at our project website

Acknowledgements