Christian Schlarmann

Adversarially Robust CLIP Models Can Induce Better (Robust) Perceptual Metrics Measuring perceptual similarity is a key tool in computer vision. In recent years perceptual metrics based on features extracted from neural networks with large and diverse training sets, e.g. CLIP, h...

📜Paper: arxiv.org/abs/2502.11725

📢 Robustness is not always at odds with accuracy! We show that adversarially robust vision encoders improve clean and robust accuracy over their base models in perceptual similarity tasks. Looking forward to presenting at SaTML @satml.org in Copenhagen next week 🇩🇰

Evaluating the Robustness of the "Ensemble Everything Everywhere" Defense Ensemble everything everywhere is a defense to adversarial examples that was recently proposed to make image classifiers robust. This defense works by ensembling a model's intermediate representations...

In line with previous works, this shows that it is important to develop adaptive attacks against new defenses in order to claim robustness.
📜 arxiv.org/abs/2411.14834

📢 Check out our new report: we show that a recently proposed defense against adversarial attacks is not robust. We circumvent gradient masking issues of the proposed model by attacking a slightly adapted surrogate model and then transferring the perturbations.

Great milestone for www.scholar-inbox.com! 🎊