Fuzzing-Based Hard-Label Black-Box Attacks Against Machine Learning Models

Anonymous

Sep 25, 2019 ICLR 2020 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: Machine learning models are known to be vulnerable to adversarial examples. Based on different levels of knowledge that attackers have about the models, adversarial example generation methods can be categorized into white-box and black-box attacks. We study the most realistic attacks, hard-label black-box attacks, where attackers only have the query access of a model and only the final predicted labels are available. The main limitation of the existing hard-label black-box attacks is that they need a large number of model queries, making them inefficient and even infeasible in practice. Inspired by the very successful fuzz testing approach in traditional software testing and computer security domains, we propose fuzzing-based hard-label black-box attacks against machine learning models. We design an AdvFuzzer to explore multiple paths between a source image and a guidance image, and design a LocalFuzzer to explore the nearby space around a given input for identifying potential adversarial examples. We demonstrate that our fuzzing attacks are feasible and effective in generating successful adversarial examples with significantly reduced number of model queries and L0 distance. More interestingly, supplied with a successful adversarial example as a seed, LocalFuzzer can immediately generate more successful adversarial examples even with smaller L2 distance from the source example, indicating that LocalFuzzer itself can be an independent and useful tool to augment many adversarial example generation algorithms.
0 Replies

Loading