Answers are grounded in this paper and cite their sources.
Purpose of the study: development of a signal-based method for determining the state of an adversarial attack on graphic objects processed by a neural network during classification. Methods of research: the use of explainable artificial intelligence (Grad-CAM) for visualization of adversarial FGSM attack is employed. A signal-based method relying on integral calculation of the gradient surface of the explanation heatmap has been proposed. Result(s): methods of explainable artificial intelligence for identifying important features and ways to utilize the obtained heatmaps for attack detection have been described and tested. In the practical part, one of the modern approaches is considered: analysis of bias and blurring of explanations using Grad-CAM. An analysis and review of the effectiveness in enhancing model resistance to attacks has been conducted. Various effects of impact resulting from attacks on attention zones and the nature of their changes have been identified. An integral method for calculating the presence of an adversarial attack in the input image has been proposed, which is applicable for automatic attack detection. Scientific novelty: the research is aimed at enhancing the informativeness regarding the nature of the attack, the degree of impact on the attacked input image, and the development of a signal-based detection method for identifying the presence of an adversarial attack.
No takes yet. Share an insight, caveat, or question.
Chekolaev et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: