Abstract
Adversarial attacks expose the inherent fragility of deep learning models, yet evaluating defenses remains challenging due to significant dataset bias. This paper identifies a specific class of "defense-friendly" images—samples that are naturally resilient to perturbations and easier to rectify using standard defensive techniques. We introduce the ImageNet-R dataset and three quantitative metrics (ARD, AMP, and ADF) to measure image robustness. Our findings demonstrate that robustness is a model-agnostic property largely determined by low-level texture features. Failing to account for these data-intrinsic properties can lead to overestimating the efficacy of adversarial defenses.
Introduction
Deep learning models are notoriously susceptible to adversarial perturbations—minimal, often imperceptible noise that triggers misclassification. While the research community has developed numerous defenses, their evaluation is often inconsistent. The authors argue that current benchmarks suffer from dataset bias, where certain images are inherently harder to fool or easier to defend. Without standardized metrics to quantify this "perturbation difficulty," researchers may inadvertently report inflated performance by testing on "easy" subsets of data.
Methods and Metrics
The researchers propose a multi-pronged approach to quantify dataset difficulty through three new metrics:
- Adversarial Robust Dataset (ARD) Score: The proportion of images in a dataset that remain correctly classified under a specific attack.
- Adversarial Minimum Perturbation (AMP) Score: The minimum amount of noise required to reach a target error rate.
- Adversarial Defense-friendly (ADF) Score: The proportion of images that successfully recover their correct label after a defense is applied.
To understand the underlying causes of robustness, the authors analyzed image textures using Gray Level Co-occurrence Matrices (GLCM). They found that robustness is highly correlated with low-level image statistics; specifically, images with lower "dissimilarity" in their texture tend to be more robust.
The ImageNet-R Dataset
The primary contribution of this work is ImageNet-R, a new benchmark consisting of over 15,000 "defense-friendly" images curated from Flickr. These images were filtered to ensure they are easy for multiple models to classify correctly and difficult for standard attacks, such as Projected Gradient Descent (PGD), to break.
Key Insights and Results
- Robustness is Model-Agnostic: If an image is difficult to perturb for a ResNet model, it is likely also difficult for a DenseNet or VGG model, suggesting robustness is a property of the data itself.
- The "Easy Image" Trap: A dataset with a high proportion of robust images gives a misleading impression of security. On ImageNet-R, models often maintain near 100% accuracy under attacks that would normally reduce them to 0% accuracy on standard datasets.
- Predictive Power of Texture: A Random Forest proxy model, using only 72 texture features from grayscale versions of images, achieved 75% accuracy in predicting whether an image would be robust. This confirms that robustness is "baked into" the image's visual structure.
Conclusion
By providing ImageNet-R and associated metrics, the authors offer the community tools to account for data robustness when evaluating AI security. This framework ensures that advancements in adversarial defense are measured against the inherent difficulty of the data, leading to more reliable and unbiased benchmarking.
Summary updated