New concolic testing method enhances Transformer robustness analysis

Researchers have developed a new concolic testing method for Transformer classifiers that uses SHAP estimates to prioritize path predicates based on their influence on the model's predictions. This approach, implemented in Python, makes self-attention semantics compatible with satisfiability modulo theories solvers. Evaluations on CIFAR-10 with compact Transformer models, ResNet18, and VGG16 demonstrated a 60% success rate in finding adversarial examples within a one-pixel budget and 900-second horizon, significantly outperforming a black-box differential evolution baseline. AI

IMPACT Enhances practical methods for finding adversarial examples in Transformer models, improving AI safety research.

RANK_REASON The cluster contains a research paper detailing a new method for testing neural network robustness. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• Chih-Duo Hong
• CIFAR-10
• Hugging Face
• Python
• ResNet18
• satisfiability modulo theories
• Shap
• transformer
• Vgg16

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →