Divide and conquer: targeted adversary detection using proximity and dependency
Date
2021
Authors
Lu, S.
Liu, L.
Li, J.
Le, T.D.
Liu, J.
Editors
Gong, Z.
Li, X.
Oguducu, S.G.
Chen, L.
Manjon, B.F.
Wu, X.
Li, X.
Oguducu, S.G.
Chen, L.
Manjon, B.F.
Wu, X.
Advisors
Journal Title
Journal ISSN
Volume Title
Type:
Conference paper
Citation
Proceedings - 12th IEEE International Conference on Big Knowledge, ICBK 2021, 2021 / Gong, Z., Li, X., Oguducu, S.G., Chen, L., Manjon, B.F., Wu, X. (ed./s), pp.125-132
Statement of Responsibility
Conference Name
12th IEEE International Conference on Big Knowledge (7 Dec 2021 - 8 Dec 2021 : Virtual)
Abstract
Deep neural networks have demonstrated impressive performance on image classification tasks, but research shows that they are susceptible to adversarial attacks. In this paper, we propose an adversary detection method, TAPD, which divides and conquers adversary detection through utilizing dependency-based and proximity-based anomaly detection techniques. We innovatively use logits to categorize adversarial examples so that the knowledge of adversary generation is captured in the generated categories. Based on the characteristics of categories, two kNN distance-based detectors are proposed, and one existing dependency-based detector is selected to effectively detect adversarial examples in their corresponding category. Extensive experiments have been conducted, and the results have shown that TAPD outperforms the state-of-the-art adversary detection methods.
School/Discipline
Dissertation Note
Provenance
Description
Access Status
Rights
Copyright 2021 The Institute of Electrical and Electronics Engineers, Inc.