Small-Perturbation Adversarial Example Detection via Proactive Decision Boundary Bending

Document Type

Article

Publication Date

1-1-2026

Abstract

Deep neural networks (DNNs) are highly susceptible to adversarial example (AE) attacks, posing serious security threats. As a specific class of adversarial attacks, small perturbation adversarial (SPA) attacks have emerged as a more severe threat to DNNs due to their subtle perturbation magnitudes, making them harder to detect. However, there is still a lack of specific solutions that are designed to defend against SPA attacks. To fill in this gap, we propose Proactive Decision Boundary Bending (PDBB) for SPA example detection in this paper. Via proactively bending the decision boundary of the target DNN, PDBB can accurately distinguish between SPA examples and benign examples. Specifically, PDBB leverages the example-perturbation training strategy to train the target DNN, enabling it to possess desirable bended decision boundaries. In addition, PDBB develops the projection-based auxiliary measurement method that accurately measures the concavity convexity of the decision boundary in the vicinity of a given test example, thereby enabling PDBB to achieve high detection accuracy. Extensive experiments demonstrate that PDBB not only achieves the state-of-the-art (SOTA) performance in detecting SPA examples but also achieves high performance in detecting general AEs (that may have relatively larger perturbations).

Publication Title

IEEE Transactions on Dependable and Secure Computing

Link to Full Text

Share

COinS