A Theoretically Grounded Extension of Universal Attacks from the Attacker's Viewpoint - Archive ouverte HAL Access content directly
Conference Papers Year : 2024

A Theoretically Grounded Extension of Universal Attacks from the Attacker's Viewpoint

Abstract

We extend universal attacks by jointly learning a set of perturbations to choose from to maximize the chance of attacking deep neural network models. Specifically, we embrace the attacker's perspective and introduce a theoretical bound quantifying how much the universal perturbations are able to fool a given model on unseen examples. An extension to assert the transferability of universal attacks is also provided. To learn such perturbations, we devise an algorithmic solution with convergence guarantees under Lipschitz continuity assumptions. Moreover, we demonstrate how it can improve the performance of state-of-the-art gradient-based universal perturbation. As evidenced by our experiments, these novel universal perturbations result in more interpretable, diverse, and transferable attacks.
Fichier principal
Vignette du fichier
ECML_2024___Semi_Universal_Perturbations-1.pdf (822.84 Ko) Télécharger le fichier
Origin Files produced by the author(s)
Licence

Dates and versions

hal-03615461 , version 1 (21-03-2022)
hal-03615461 , version 2 (07-06-2023)
hal-03615461 , version 3 (07-06-2024)

Licence

Identifiers

  • HAL Id : hal-03615461 , version 3

Cite

Jordan Patracone, Paul Viallard, Emilie Morvant, Gilles Gasso, Amaury Habrard, et al.. A Theoretically Grounded Extension of Universal Attacks from the Attacker's Viewpoint. European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, Sep 2024, Vilnius, Lithuania. ⟨hal-03615461v3⟩
133 View
174 Download

Share

Gmail Mastodon Facebook X LinkedIn More