Institutional Repository of Vilnius University
Priešiškų atakų įtaka paaiškinamiesiems dirbtinio intelekto modeliams /
Abstract
dc:descriptionThis paper presents the impact of adversarial attacks on explainable artificial intelligence methods. It focuses on a specific group of attacks — adversarial examples — and posthoc XAI methods. The study presents how the images affected by attacks were generated and how the XAI methods were evaluated. Image similarity metrics were used to assess the explanations, and a user study was conducted where a group of human evaluators rated the explanations. According to the results, the bestperforming methods were GradCAM++ and GradCAM. The worstperforming methods were LRP and Guided GradCAM. The most effective attack across all methods was FGSM, although it was also the most easily noticeable by humans. The attacks that distorted the explanations the least were Square Attack (l∞) and Square Attack (l2). Feature Adversaries was the second most effective attack across all methods and the hardest to detect for humans.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Pakenis, Armintas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD239031979&prefLang=en_US