Think you can spot the difference? The AI can't. Panda or Gibbon? A Beginner's Introduction to Adversarial Attacks is an interactive, beginner-friendly visualization that introduces how machine-learning models can be fooled by malicious adversarial attacks. Built primarily with D3.js and Idyll, the guide focuses on the Fast Gradient Sign Method (FGSM) and shows how tiny, human-imperceptible tweaks to an image can push a ResNet-34 model into making confident mistakes. With dynamic visuals and animations, users can compare clean and subtly perturbed images, explore how these attacks shift model behavior, and examine two versions of ResNet-34, one trained normally and one trained with adversarial methods, to see how they respond differently.
- Links:
- Explainable: https://visxai-aml.vercel.app/
- Video Demo: https://youtu.be/ASEd4f5gMvA
- Recognition & Outreach
- Accepted and presented at the 7th VISxAI Workshop at IEEE VIS24: VISxAI Workshop Program Info
Presentation at the VISxAI Workshop
- Core Features
- Explains adversarial attacks using beginner-friendly interactive visualizations.
- Explores the FGSM attack's impact on ResNet-34 models, with insights into both natural and adversarial images, as well as standard and adversarial trainings.
- Includes embedding-level and instance-level analysis to show how adversarial perturbations affect models.
- SkillsPython, PyTorch, t-SNE, Adversarial Machine Learning, XAI Visualization, D3.js, Idyll-lang
- AuthorsYuzhe You, Jian Zhao
- KeywordsAdversarial Machine Learning, FGSM Attack, Adversarial Attack, Image Classification, Visualization, ResNet, Model Robustness