SCOOTER: A Human Evaluation Framework for Unrestricted Adversarial Examples

Unrestricted adversarial attacks aim to fool computer vision models without being constrained by $\ell_p$-norm bounds to remain imperceptible to humans, for example, by changing an object's color. This allows attackers to circumvent traditional, norm-bounded defense strategies such as adversarial training or certified defense strategies. However, due to their unrestricted nature, there are also no guarantees of norm-based imperceptibility, necessitating human evaluations to verify just how authentic these adversarial examples look. While some related work assesses this vital quality of adversarial attacks, none provide statistically significant insights. This issue necessitates a unified framework that supports and streamlines such an assessment for evaluating and comparing unrestricted attacks. To close this gap, we introduce SCOOTER - an open-source, statistically powered framework for evaluating unrestricted adversarial examples. Our contributions are: $(i)$ best-practice guidelines for crowd-study power, compensation, and Likert equivalence bounds to measure imperceptibility; $(ii)$ the first large-scale human vs. model comparison across 346 human participants showing that three color-space attacks and three diffusion-based attacks fail to produce imperceptible images. Furthermore, we found that GPT-4o can serve as a preliminary test for imperceptibility, but it only consistently detects adversarial examples for four out of six tested attacks; $(iii)$ open-source software tools, including a browser-based task template to collect annotations and analysis scripts in Python and R; $(iv)$ an ImageNet-derived benchmark dataset containing 3K real images, 7K adversarial examples, and over 34K human ratings. Our findings demonstrate that automated vision systems do not align with human perception, reinforcing the need for a ground-truth SCOOTER benchmark.

How Real Is Real? A Human Evaluation Framework for Unrestricted Adversarial Examples

With an ever-increasing reliance on machine learning (ML) models in the real world, adversarial examples threaten the safety of AI-based systems such as autonomous vehicles. In the image domain, they represent maliciously perturbed data points that look benign to humans (i.e., the image modification is not noticeable) but greatly mislead state-of-the-art ML models. Previously, researchers ensured the imperceptibility of their altered data points by restricting perturbations via $\ell_p$ norms. However, recent publications claim that creating natural-looking adversarial examples without such restrictions is also possible. With much more freedom to instill malicious information into data, these unrestricted adversarial examples can potentially overcome traditional defense strategies as they are not constrained by the limitations or patterns these defenses typically recognize and mitigate. This allows attackers to operate outside of expected threat models. However, surveying existing image-based methods, we noticed a need for more human evaluations of the proposed image modifications. Based on existing human-assessment frameworks for image generation quality, we propose SCOOTER - an evaluation framework for unrestricted image-based attacks. It provides researchers with guidelines for conducting statistically significant human experiments, standardized questions, and a ready-to-use implementation. We propose a framework that allows researchers to analyze how imperceptible their unrestricted attacks truly are.