Research map / Adversarial robustness and security / Certified and empirical robustness
Extending Adversarial Attacks to Produce Adversarial Class Probability Distributions
Jon Vadillo, Roberto Santana, Jose A. Lozano
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use informed transition matrices (EWTM) instead of AM | target distribution approximation · Speech Commands, DeepFool | max/mean abs diff, KL, Spearman | Improved |
| use MAB baseline instead of proposed AM/UBM(eps>0.01)/CRM | fooling rate · Speech Commands, DeepFool, eps=0.15 | fooling rate (%): 94.69 -> 79.64 | Hurt |
| use EWTM instead of AM | fooling rate · Speech Commands, DeepFool, eps=0.15 | fooling rate (%): 94.69 -> 87.84 | Hurt |
| add upper-bound relaxation via eta to UBM (ablation) | UBM · Speech Commands, DeepFool, eps=0.01 | valid matrix success (%): 67.9 -> 100.0 | Improved |
| add Laplace smoothing to CRM (ablation) | CRM · Speech Commands, DeepFool, eps=0.05 | valid matrix success (%): 14.0 -> 70.6 | Improved |
| add zeroing of P({y_i}|y_i) to CRM (ablation) | CRM · Speech Commands, DeepFool, eps=0.05 | valid matrix success (%): 70.6 -> 100.0 | Improved |
| increase maximum distortion eps (ablation) | all methods · Speech Commands, DeepFool | target approximation: max abs diff below 0.09, mean abs diff 0.03, Spearman above 0.80 for eps >= 0.05 | Improved |
| use EWTM instead of other methods at high distortion | EWTM · Speech Commands, DeepFool, eps>=0.05 | distribution similarity | Improved |
| increase eps when reproducing the original distribution (ablation) | all methods · Speech Commands, DeepFool, original distribution | approximation accuracy | Hurt |
| reduce samples per class N to 1 (ablation) | EWTM · Speech Commands, DeepFool, eps=0.01 | valid matrix success (%): 100 -> 0.8 | Hurt |
| reduce samples per class N to 50 (ablation) | AM/UBM/EWTM · Speech Commands, DeepFool, eps=0.01 | effectiveness | No clear effect |
| use sparse-reach attack FGSM instead of C&W/PGD | proposed methods · Speech Commands, DeepFool | differences between methods | Mixed |
Closest papers by meaning
- Adversarial Defense via Learning to Generate Diverse Attacks (ICCV 2019 · similarity 0.88)
- Statistical Estimation of Adversarial Risk in Large Language Models under Best-of-N Sampling (NeurIPS 2026 · similarity 0.87)
- Confidence Elicitation: A New Attack Vector for Large Language Models (ICLR 2025 · similarity 0.87)
- Strong Transferable Adversarial Attacks via Ensembled Asymptotically Normal Distribution Learning (CVPR 2024 · similarity 0.86)
- AttackBench: Evaluating Gradient-based Attacks for Adversarial Examples (AAAI 2025 · similarity 0.86)
- Understanding the Robustness of Randomized Feature Defense Against Query-Based Adversarial Attacks (ICLR 2024 · similarity 0.86)
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.
