"Exploring Comparative Visual Approaches for Understanding Model Trade-offs in Adversarial Machine Learning" is a project that explores effective comparative visual techniques for model trade-off analysis in adversarial machine learning (AML). This work focuses on the need for practitioners to concurrently evaluate the accuracy and robustness of AI models across both standard and adversarial conditions. The project identifies and presents effective visualization techniques for understanding AML trade-offs at different levels of detail. Based on insights gathered from an iterative design process, we designed VATRA, a design probe that supports multi-level comparative visualizations and enables machine learning practitioners to analyze trade-offs in adversarial training (AT) methods, which often come at the cost of reduced model accuracy and fairness.

Recognition & Outreach
• In submission to CHI 2025.

Core Features
• Exploration of comparative visual designs for model trade-off analysis in standard and adversarial conditions.
• Multi-level comparative visual analytics: from instance-level, embedding-level, to model-level comparisons.
• Empirical insights into design considerations for using comparative visualizations in AML trade-off analysis.

• SkillsPython, PyTorch, Grad-CAM, Visualization Design, Machine Learning, Evasion Attacks, D3.js, JavaScript
• AuthorsYuzhe You, Jian Zhao
• KeywordsAdversarial Machine Learning, Adversarial Training, Visualization, Model Robustness, Accuracy-Robustness Trade-offs