Understanding the Limitations of Pixel Discretization Against Adversarial Attacks
Towards Understanding Limitations of Pixel Discretization Against Adversarial Attacks
Translation in progress — untranslated text is omitted.
The researchers analyzed pixel discretization, a preprocessing defense against adversarial attacks, and more complex variants that account for dataset characteristics; the results show that their resistance to strong white-box attacks remains poor. Such methods are attractive in practice because they are simple, computationally cheap, and applicable to many systems; they had previously been observed to be effective on simple datasets such as MNIST, but on complex datasets such as ImageNet they are broken by the newly proposed strong white-box attacks. The authors analyze these results in a theoretical framework, giving strong evidence that pixel discretization is unlikely to be effective except on the simplest datasets, and their argument also suggests why some other preprocessing defenses may be unsafe.
- Preprocessing defenses such as pixel discretization are simple and practical, effective on MNIST, but fail on complex datasets under strong white-box BPDA attacks; why they fail and whether they can be improved remains unclear.
- The authors view discretization as clustering of pixels, propose two data-dependent codeword construction algorithms, density estimation and k-medoids, and establish a theoretical framework: under an ideal generative model they prove robustness can be improved, and they also give a computable certified lower bound and analyses of equivalence classes and separability.
- On MNIST, discretization raises the adversarially trained model's robustness to 97.32% (2-bin, Table I), and gives a certified lower bound of 96.43% at ε=0.1; on CIFAR-10, GTSRB, and ImageNet the improvement over the undefended baseline is very small, which the authors attribute to the lack of separable clusters in complex data.
- The authors state that the analysis is limited to pixel-level discretization, and leave generalizing the intuition to more general preprocessing defenses as future work (Section VI); the experiments cover only ℓ∞ attacks, five image datasets and two codeword construction algorithms, and certification is computed only on MNIST.