Detecting Adversarial Examples from Sensitivity Inconsistency of Spatial-Transform Domain
Mahalanobis distance
Decision boundary
Deep Neural Networks
Feature vector
DOI:
10.1609/aaai.v35i11.17187
Publication Date:
2022-09-08T19:23:47Z
AUTHORS (4)
ABSTRACT
Deep neural networks (DNNs) have been shown to be vulnerable against adversarial examples (AEs), which are maliciously designed cause dramatic model output errors. In this work, we reveal that normal (NEs) insensitive the fluctuations occurring at highly-curved region of decision boundary, while AEs typically over one single domain (mostly spatial domain) exhibit exorbitant sensitivity on such fluctuations. This phenomenon motivates us design another classifier (called dual classifier) with transformed can collaboratively used original primal detect AEs, by virtue inconsistency. When comparing state-of-the-art algorithms based Local Intrinsic Dimensionality (LID), Mahalanobis Distance (MD), and Feature Squeezing (FS), our proposed Sensitivity Inconsistency Detector (SID) achieves improved AE detection performance superior generalization capabilities, especially in challenging cases where perturbation levels small. Intensive experimental results ResNet VGG validate superiority SID.
SUPPLEMENTAL MATERIAL
Coming soon ....
REFERENCES (0)
CITATIONS (30)
EXTERNAL LINKS
PlumX Metrics
RECOMMENDATIONS
FAIR ASSESSMENT
Coming soon ....
JUPYTER LAB
Coming soon ....