SemiAdv: Query-Efficient Black-Box Adversarial Attacks Based on Unlabeled Images

SemiAdv introduces semi-supervised learning into black-box adversarial attacks, requiring on average only hundreds of queries to achieve an effective attack with over 90% success rate. Existing black-box attacks typically require large amounts of labeled data to bridge the gap between the substitute model and the target model, whereas SemiAdv can directly leverage unlabeled raw data to launch attacks. Experiments show that, compared with state-of-the-art attacks, it saves up to 12 times the number of queries when generating adversarial examples, while the attack success rate remains competitive.