Boosting Transferable Attacks with Decision-based Adversarial Perturbations

Deep Neural Networks (DNNs) are vulnerable to carefully crafted adversarial examples, which indicates that DNNs have frangibility that can be easily exploited in security-critical scenarios. Investigating adversarial strategies is essential for developing corresponding adversarial defenses and building DNNs to effectively deal with such challenges. However, most existing adversarial frameworks focused on improving the adversarial performance via salient feature argumentation ignoring the decision-based patterns derived from deep neural task models, which may limit the adversarial generalization of adversarial examples in invading DNNs. To this end, we introduce a simple yet flexible adversarial perturbation generation framework named Decision-based Adversarial Method (DAM), which first calls the logit-based decisions from both original inputs and crafted entities, then adaptively correlates such generated decision-based patterns with general gradients following an adaptive normalized manner to form effective adversarial characteristics. Extensive experiments demonstrate that adversarial features generated by DAM achieve extraordinary generalization capability in misleading deep neural networks under complex constraints while preserving adaptable compatibility with existing adversarial frameworks.