Exploring Adversarial Attacks and Defenses in Vision Transformers trained with DINO

TL;DR

This paper analyzes the robustness of self-supervised Vision Transformers trained with DINO against adversarial attacks, exploring latent space properties and evaluating defense strategies like adversarial training and ensemble methods.

Contribution

It provides the first analysis of adversarial robustness in DINO-trained Vision Transformers and assesses defense strategies with limited fine-tuning.

Findings

Self-supervised features show different robustness properties than supervised ones.

Latent space attacks reveal specific vulnerabilities.

Defense strategies improve robustness with minimal fine-tuning.

Abstract

This work conducts the first analysis on the robustness against adversarial attacks on self-supervised Vision Transformers trained using DINO. First, we evaluate whether features learned through self-supervision are more robust to adversarial attacks than those emerging from supervised learning. Then, we present properties arising for attacks in the latent space. Finally, we evaluate whether three well-known defense strategies can increase adversarial robustness in downstream tasks by only fine-tuning the classification head to provide robustness even in view of limited compute resources. These defense strategies are: Adversarial Training, Ensemble Adversarial Training and Ensemble of Specialized Networks.