Research map / Adversarial robustness and security / Certified and empirical robustness
Towards Certifiably Robust Face Recognition
Seunghun Paik, Dongsoo Kim, Chanwoo Hwang, Sunpill Kim, Jae Hong Seo
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 |
|---|---|---|---|
| add feature-norm regularization loss (xi=3) to a 1-Lipschitz face recognition model (ablation) | 1-Lipschitz FR model (SLL + ElasticFace) · LFW untargeted certification eps=1.5 | CRA_U (%): 0 -> 100 | Improved |
| increase regularization strength xi from 0 to 5 (ablation) | 1-Lipschitz FR model · LFW | TAR@FAR=1e-3 (%): 97.73 -> 65.27 | Hurt |
| increase regularization strength xi from 0 to 5 (ablation) | 1-Lipschitz FR model · LFW impersonation certification eps=0.75 | CRA_I (%): 0 -> 93.95 | Improved |
| add feature-norm regularization (xi=3) vs xi=0 for empirical robustness (ablation) | 1-Lipschitz FR model · LFW untargeted PGD20 eps=3.0 | ERA_U (%): 0.18 -> 100 | Improved |
| increase xi from 0 to 5 for dodging certification at eps=0.25 (ablation) | 1-Lipschitz FR model · LFW dodging certification eps=0.25 | CRA_D (%): 0 -> 46.68 (xi=1) -> 66.82 (xi=3) -> 53.06 (xi=5) | Mixed |
| use certified 1-Lipschitz model instead of non-robust ArcFace/ElasticFace/AdaCos | Face recognition · LFW untargeted PGD20 eps=3.0 | ERA_U (%): 0.22 -> 100 (vs ArcFace) | Improved |
| use certified 1-Lipschitz model on small-eps dodging attack | Face recognition · LFW dodging PGD20 eps=0.25 | ERA_D (%): 98.03 (ArcFace) vs 74.22 (Ours) | Hurt |
| use 1-Lipschitz backbone instead of non-robust backbone | Face recognition · LFW | TAR@FAR=1e-3 (%): 99.75 (ArcFace) vs 83.97 (Ours) | Hurt |
| use classification-style header (LLN plus margin) instead of ElasticFace for 1-Lipschitz FR (ablation) | 1-Lipschitz FR model · Face recognition training | accuracy | Hurt |
| compare with adversarially trained FR models (PGD-AT, TRADES) | Face recognition · LFW | TAR@FAR=1e-3 (%): 28.47-67.17 vs 83.97 (ours) | Hurt |
Closest papers by meaning
- Lower bounds on the robustness to adversarial perturbations (NeurIPS 2017 · similarity 0.89)
- Detecting and Mitigating Adversarial Perturbations for Robust Face Recognition (IJCV 2019 · similarity 0.89)
- A Recipe for Improved Certifiable Robustness (ICLR 2024 · similarity 0.88)
- Lipschitz-Margin Training: Scalable Certification of Perturbation Invariance for Deep Neural Networks (NeurIPS 2018 · similarity 0.88)
- Learning Better Certified Models from Empirically-Robust Teachers (NeurIPS 2026 · similarity 0.87)
- Probably Approximately Global Robustness Certification (ICML 2025 · similarity 0.87)
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.
