Introduction
In the evolving field of adversarial machine learning, researchers frequently report high accuracy for new defense mechanisms against sophisticated attacks. However, these results are often skewed by an overlooked factor: dataset bias. Not all images are equally susceptible to adversarial perturbations. Some images possess an inherent "natural robustness," requiring significantly more noise to induce misclassification and proving much easier for denoising algorithms to rectify.
This paper identifies a specific class of "defense-friendly" images that are naturally resilient to adversarial attacks. To prevent over-optimistic reporting in the field, we introduce ImageNet-R—a dataset of over 15,000 robust images—and propose three new metrics to quantify the inherent difficulty of defending a given dataset.
Quantifying Data Robustness
To move beyond anecdotal evidence of image difficulty, we propose a multi-pronged approach to identifying and measuring image robustness through three primary scoring mechanisms:
- Adversarial Robust Dataset (ARD) Score: The proportion of images in a dataset that resist a specific attack (e.g., PGD) across multiple architectures.
- Adversarial Minimum Perturbation (AMP) Score: The minimum noise threshold required to successfully "break" a specific percentage of the dataset.
- Adversarial Defense-friendly (ADF) Score: The proportion of images that a defense can successfully recover after they have been subjected to an attack.
Categorizing Image Difficulty
Our research categorizes images into three distinct tiers based on their interaction with attacks and defenses. "Easy" images are those correctly classified by clean models; "$ε$-robust" images resist perturbations within a specific bound; and "Defense-friendly" images are those that, even if perturbed, are easily restored by standard defense layers.
Key Insights: Robustness vs. Confidence
A common misconception in adversarial research is that high-confidence predictions (indicated by a large "logit gap" between the top two classes) imply higher robustness. Our analysis debunked this assumption. We found that many non-robust images exhibit high confidence scores, while some robust images do not.
Instead of confidence, we discovered that texture is a primary indicator of robustness. Using Grey Level Co-occurrence Matrices (GLCM) to extract statistical features, we found that images with lower "dissimilarity" (smoother, more consistent textures) are naturally more robust. We successfully trained a Random Forest classifier to predict image robustness with 75% accuracy based solely on these texture features, providing a low-cost proxy for measuring dataset difficulty.
Experimental Results and Benchmarking
Our evaluation of standard benchmarks revealed a significant performance gap. On specifically curated "Non-Robust" datasets, model accuracy plummeted to 0% under PGD attacks. Conversely, on our "Robust" (R1) dataset, models like ResNet and DenseNet maintained near 100% accuracy under identical attack parameters.
Furthermore, when applying the BaRT defense, robust images achieved recovery rates (ADF scores) as high as 0.98, while non-robust images often remained below 0.15. This confirms that evaluating a defense on a dataset with a high ARD score provides a misleading impression of the defense's actual utility.
Conclusion
By identifying defense-friendly images and providing the ImageNet-R dataset, this work establishes a more rigorous standard for adversarial benchmarking. Researchers must account for the inherent robustness of their evaluation data to ensure that reported improvements in defense accuracy are due to the algorithm itself rather than favorable data characteristics. Future work will explore whether these robust texture features can be synthetically engineered into sensitive images to enhance their natural resilience.
Summary updated