Research map / Architectures and efficiency / Attention and transformers
Robustness Verification for Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace IBP bounds with the proposed linear-relaxation verification method | 1-layer Transformer · Yelp, 1-position ℓ₂ perturbation | certified lower bound (Min): 1.4E-4 -> 0.384 | Helped |
| replace IBP bounds with the proposed linear-relaxation verification method | 1-layer Transformer · SST, 1-position ℓ₂ perturbation | certified lower bound (Min): 2.4E-4 -> 0.418 | Helped |
| replace IBP bounds with the proposed linear-relaxation verification method | 1-layer Transformer · Yelp, 2-position ℓ₂ perturbation | certified lower bound (Min): 6.5E-5 -> 0.242 | Helped |
| replace IBP bounds with the proposed linear-relaxation verification method | 1-layer Transformer · SST, 2-position ℓ₂ perturbation | certified lower bound (Min): 1.1E-4 -> 0.212 | Helped |
| use fully-forward propagation instead of backward and forward propagation (ablation) | Transformer verification · Yelp, 1-position ℓ₂ perturbation | certified lower bound (Avg): 0.946 -> 0.599 | Hurt |
| use fully-forward propagation instead of backward and forward propagation (ablation) | Transformer verification · SST, 1-position ℓ₂ perturbation | certified lower bound (Avg): 0.439 -> 0.366 | Hurt |
| use backward and forward propagation instead of fully-backward propagation (ablation) | Transformer verification · Yelp, 1-position ℓ₂ perturbation | certified lower bound (Avg): 0.947 -> 0.946 | No clear effect |
| use backward and forward propagation instead of fully-backward propagation (ablation) | Transformer verification · Yelp, 1-position ℓ₂ perturbation | computation time (sec): 140.4 -> 26.0 | Helped |
| use backward and forward propagation instead of fully-backward propagation (ablation) | Transformer verification · SST, 1-position ℓ₂ perturbation | computation time (sec): 178.8 -> 24.3 | Helped |
| use certified lower bounds rather than gradient or upper-bound estimates to rank important words | 1-layer Transformer · SST word importance identification | importance score of identified most important words: 0.47 / 0.45 -> 0.57 | Helped |
| use certified lower bounds rather than gradient or upper-bound estimates to rank least important words | 1-layer Transformer · SST word importance identification | importance score of identified least important words: 0.40 / 0.24 -> 0.01 | Helped |
Closest papers by meaning
- GaLileo: General Linear Relaxation Framework for Tightening Robustness Certification of Transformers (AAAI 2024 · similarity 0.88)
- Transformers Learn Low Sensitivity Functions: Investigations and Implications (ICLR 2025 · similarity 0.84)
- Probably Approximately Global Robustness Certification (ICML 2025 · similarity 0.84)
- Interpreting Robustness Proofs of Deep Neural Networks (ICLR 2024 · similarity 0.83)
- ProTransformer: Robustify Transformers via Plug-and-Play Paradigm (NeurIPS 2024 · similarity 0.83)
- Adventures of Trustworthy Vision-Language Models: A Survey (AAAI 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.