Reproduce “Adversarial Robustness of Discriminative Self-Supervised Learning in Vision” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ömer Veysel Çağatan, Ömer Faruk Tal, M. Emre Gürsoy

Koç University researchers conducted a systematic evaluation of adversarial robustness across various discriminative self-supervised learning (SSL) models and a supervised baseline in vision tasks. The study found that SSL models generally exhibit better robustness on ImageNet classification and linear transfer learning, but this does not consistently transfer to all downstream tasks, with architectural choices and algorithmic aspects significantly influencing resilience.

Key insights

  • Discriminative SSL models generally demonstrate improved adversarial robustness on ImageNet classification compared to supervised counterparts, with specific architectural components and momentum-based learning contributing to resilience.
  • Adversarial robustness established during pretraining transfers effectively to linear evaluation on downstream tasks, but fine-tuning significantly reshapes defenses against instance-specific attacks while susceptibility to universal perturbations persists.
  • The impact of model architecture on adversarial robustness is highly dependent on the SSL objective, with DINO showing improved robustness with Vision Transformers, while MoCoV3's performance decreased with ViT backbones.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.