Using features from an adversarially robust ImageNet model with a k-nearest-neighbor score gives state-of-the-art adversarial robustness in novelty detection on several image benchmarks.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Killing it with Zero-Shot: Adversarially Robust Novelty Detection
Using features from an adversarially robust ImageNet model with a k-nearest-neighbor score gives state-of-the-art adversarial robustness in novelty detection on several image benchmarks.