IC-077VGG-16's loss landscape barrier height distinguishes adversarial from real inputs, enabling a detection method that outperforms baselines on DeepFool and C&W attacks

Jakub Vrabel, Ori Shem-Ur, Yaron Oz, David Krueger

SourceInput Space Mode Connectivity in Deep Neural Networks

Using VGG-16 on CIFAR-10/100, the authors compute loss curves along linear interpolant paths between each test image and a class reference image, concatenate these with sorted logits, and train a KNN classifier. The method exploits the fact that real-adversarial pairs exhibit higher and more complex loss barriers than real-real pairs. On CIFAR-10, the method achieves 91.4/96.9 (accuracy/AUC) on DeepFool and 93.7/98.3 on C&W, outperforming the best baselines (LayerPFS: 86.1/92.2 and 86.8/93.3). However, it underperforms on simpler attacks (FGSM: 69.8/76.8 vs LayerMFS 99.6/100) because those attacks do not create the distinct loss barrier the method relies on.

Evidence
correlational
Key metric
CIFAR-10: DeepFool 91.4/96.9, C&W 93.7/98.3 (mode connectivity) vs LayerPFS 86.1/92.2, 86.8/93.3. CIFAR-100: DeepFool 86.8/94.0, C&W 92.9/97.3. FGSM: 69.8/76.8 (underperforms LayerMFS 99.6/100).
Caveat
The method does not require access to feature maps but may fail against adaptive attacks, which were not tested due to the non-differentiability of the current detection algorithm. Underperforms on simpler attacks (FGSM, BIM, PGD) that do not produce a distinct loss barrier.
Model
VGG / VGG13 VGG-16
Datasets
CIFAR-10 [eval], CIFAR-100 [eval]
Methods
DeepFool [eval], C&W [eval], FGSM (Fast Gradient Sign Method) [eval], PGD (Projected Gradient Descent) [eval]
Related work
SpectralDefense (Harder et al. 2021) [compared-to]
Related findings
IC-076
Extraction
automatic-extraction