Linear Modeling of the Adversarial Noise Space - Inria - Institut national de recherche en sciences et technologies du numérique Accéder directement au contenu
Communication Dans Un Congrès Année : 2024

Linear Modeling of the Adversarial Noise Space

Résumé

Recent works have revealed the vulnerability of deep neural network (DNN) classifiers to adversarial attacks. Among such attacks, it is common to distinguish specific attacks adapted to each example from universal ones, referred to as example-agnostic. Even though specific adversarial attacks are efficient on their target DNN classifier to attack, they struggle to transfer to others. Conversely, universal adversarial attacks suffer from lower attack success. To reconcile universality and efficiency, we propose LIMANS, a model of the adversarial noise space, allowing to frame any specific adversarial perturbation as a linear combination of the universal adversarial directions. We bring in two stochastic gradient based algorithms for learning these universal directions and the associated adversarial attacks. Empirical analyses conducted with the CIFAR-10 and ImageNet datasets show that LIMANS (i) enables crafting specific and robust adversarial attacks with high probability, (ii) provides a deeper understanding of DNN flaws, and (iii) shows significant ability in transferability.
Fichier principal
Vignette du fichier
LIMANS_ECML_2024.pdf (1.69 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-04589704 , version 1 (27-05-2024)

Licence

Identifiants

  • HAL Id : hal-04589704 , version 1

Citer

Jordan Patracone, Lucas Anquetil, Yuan Liu, Gilles Gasso, Stéphane Canu. Linear Modeling of the Adversarial Noise Space. ECML PKDD 2024 - European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, Sep 2024, Vilnius, Lithuania. ⟨hal-04589704⟩
27 Consultations
19 Téléchargements

Partager

Gmail Mastodon Facebook X LinkedIn More