Graph-based methods coupled with specific distributional distances for adversarial attack detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Artificial neural networks (ANNs) are vulnerable to adversarial attacks, where subtle input perturbations cause misclassifications.
• Extensive research has focused on detecting and defending against these adversarial attacks.

Purpose of the Study:

• To introduce a novel graph-based approach for detecting and interpreting adversarial attacks in ANNs.
• To provide a new perspective on understanding the nature of adversarial examples.

Main Methods:

• Constructing a sparse graph from an input image using layer-wise relevance propagation, retaining only high-relevance edges.
• Computing three graph-based quantities and comparing them to those from the training set.
• Employing two classification methods: Wasserstein distance on node degree and logistic regression.

Main Results:

• The proposed graph-based interpretation successfully classifies images as benign or adversarial.
• Both classification methods demonstrated strong performance in identifying adversarial examples.
• The approach validates the utility of graph-based analysis for adversarial attack detection.

Conclusions:

• A graph-based perspective offers a valuable method for detecting and interpreting adversarial attacks on neural networks.
• The layer-wise relevance propagation and graph analysis provide a robust framework for enhancing ANN security.
• This research contributes to the ongoing effort to build more resilient AI systems against malicious manipulation.