Conference proceeding
An Information-Theoretic Explanation for the Adversarial Fragility of AI Classifiers
2019 IEEE International Symposium on Information Theory (ISIT), Vol.2019-, pp.1977-1981
07/2019
DOI: 10.1109/ISIT.2019.8849757
Abstract
We present a simple hypothesis about a compression property of artificial intelligence (AI) classifiers and present theoretical arguments to show that this hypothesis successfully accounts for the observed fragility of AI classifiers to small adversarial perturbations. We also propose a new method for detecting when small input perturbations cause classifier errors, and show theoretical guarantees for the performance of this detection method. We present experimental results with a voice recognition system to demonstrate this method. The ideas in this paper are motivated by a simple analogy between AI classifiers and the standard Shannon model of a communication system. 1
Details
- Title: Subtitle
- An Information-Theoretic Explanation for the Adversarial Fragility of AI Classifiers
- Creators
- Hui Xie - University of IowaJirong Yi - University of IowaWeiyu Xu - University of IowaRaghu Mudumbai - University of Iowa
- Resource Type
- Conference proceeding
- Publication Details
- 2019 IEEE International Symposium on Information Theory (ISIT), Vol.2019-, pp.1977-1981
- Publisher
- IEEE
- DOI
- 10.1109/ISIT.2019.8849757
- ISSN
- 2157-8095
- eISSN
- 2157-8117
- Language
- English
- Date published
- 07/2019
- Academic Unit
- Electrical and Computer Engineering
- Record Identifier
- 9984197525902771
Metrics
9 Record Views