paper:doi-10-1109-isit-2019-8849757An Information-Theoretic Explanation for the Adversarial Fragility of AI Classifiers
Original abstract (expand)
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
Similar preprints — Semantic Scholar
Cited by (1)
- Adversarial Examples Are Not Bugs, They Are Superposition
Superposition causally contributes to adversarial vulnerability, and the link runs in both directions: in a toy ReLU autoencoder (n=100 features, m=20 hidden units), raising feature sparsity raises fe