GRAPHITE: Generating Automatic Physical Examples for Machine-Learning Attacks on Computer Vision Systems
GRAPHITE: Generating Automatic Physical Examples for Machine-Learning Attacks on Computer Vision Systems
Translation in progress — untranslated text is omitted.
The researchers propose the GRAPHITE framework, which can automatically generate physical adversarial examples that can be made into stickers and attack computer vision models in black-box hard-label settings. The method is based on a transform-robustness metric that takes the expectation over transformations, automatically generates small-size masks and solves them with gradient-free optimization, and can trade off among transform robustness, perturbation size and number of queries. In the hard-label black-box setting on the GTSRB model, the authors claim to have found attacks for all 1,806 victim-target class pairs, with an average transform robustness of 77.8%, perturbations accounting for 16.63% of the victim image, and about 126K queries per pair. With digital attacks only, 92.2% of class pairs can be attacked with small-size patches using an average of 566 queries. The paper also claims that perturbations modifying a small region of the image can break the PatchGuard defense against patch attacks.
- Real-world adversarial attacks must simultaneously satisfy three points: automatically restricting the perturbation region so it can be made into a sticker, robustness to physical transformations such as viewpoint and lighting, and working under a hard-label black-box setting where only the top-1 label is available; the authors state that existing methods cannot satisfy all three simultaneously.
- The authors propose GRAPHITE: it formulates mask size and transform-robustness (EoT) as a joint optimization; the white-box version extends C&W l0 with EoT PGD; the black-box version splits into two steps: mask generation (heatmap estimation + coarse/fine-grained reduction) and RGF perturbation boosting.
- Average 77.8% transform-robustness, mask 16.63%, and 126K queries over all 1,806 victim-target pairs in GTSRB; physical sticker and drive-by test success rate up to 97.5%; average 68% TR for attacks on the PatchGuard-defended CIFAR-10 model.
- The authors point out that gamma correction struggles to simulate sunlight glare and printing errors (especially blue), and that the Turn Left Ahead physical attack has only 0% outdoor success rate; in-depth evaluation of defenses, adversarial training, and the feasibility of query detection defenses are all left for future work.