An Information-Theoretic Explanation for the Adversarial Fragility of AI Classifiers
arXiv:1901.09413
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.
arXiv comment5 pages
References in corpus (8)
- Explaining and Harnessing Adversarial Examples
- Ensemble Adversarial Training: Attacks and Defenses
- Opening the Black Box of Deep Neural Networks via Information
- Defensive Distillation is Not Robust to Adversarial Examples
- Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey
- A Boundary Tilting Persepective on the Phenomenon of Adversarial Examples
- DARCCC: Detecting Adversaries by Reconstruction from Class Conditional Capsules
- BiHMP-GAN: Bidirectional 3D Human Motion Prediction GAN