Art der Publikation: Beitrag in Sammelwerk

ExplAInable Pixels: Investigating One-Pixel Attacks on Deep Learning Models with Explainable Visualizations

Autor(en):

Keppel, Jonas; Liebers, Jonathan; Auda, Jonas; Gruenefeld, Uwe; Schneegass, Stefan

Titel des Sammelbands:

Proceedings of the 21st International Conference on Mobile and Ubiquitous Multimedia

Seiten:

231-242

Verlag:

Association for Computing Machinery

Ort(e):

New York, NY, USA

Veröffentlichung:

2022

ISBN:

9781450398206

Schlagworte:

human-in-the-loop, explainability, adversarial examples, one-pixel attacks

Digital Object Identifier (DOI):

doi:10.1145/3568444.3568469

Zitation:

Download BibTeX

Kurzfassung

Nowadays, deep learning models enable numerous safety-critical applications, such as biometric authentication, medical diagnosis support, and self-driving cars. However, previous studies have frequently demonstrated that these models are attackable through slight modifications of their inputs, so-called adversarial attacks. Hence, researchers proposed investigating examples of these attacks with explainable artificial intelligence to understand them better. In this line, we developed an expert tool to explore adversarial attacks and defenses against them. To demonstrate the capabilities of our visualization tool, we worked with the publicly available CIFAR-10 dataset and generated one-pixel attacks. After that, we conducted an online evaluation with 16 experts. We found that our tool is usable and practical, providing evidence that it can support understanding, explaining, and preventing adversarial examples.