Adversarial poisoning and inverse poisoning against deep learning
Résumé
Efficient attacks for both adversarial poisoning and adversarial inverse poisoning have recently been found on frozen deep feature plus support vector machine. But, new experiments show that such attacks only works for poisoning, but, not inverse poisoning when targeting deep networks. Investigating this observation, this paper found that adversarial poisoning attacks based on energetic landscape modification outperforms previous ones based on energetic minimum modification. This way, this paper shows that stochastic training is not sufficient defence against inverse poisoning.
Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...