Research map / Adversarial robustness and security / Certified and empirical robustness
FaceSec: A Fine-Grained Robustness Evaluation Framework for Face Recognition Systems
Liang Tong, Zhengzhang Chen, Jingchao Ni, Wei Cheng, Dongjin Song, Haifeng Chen, Yevgeniy Vorobeychik
Check 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 |
|---|---|---|---|
| Expose the target architecture rather than the training set to the attacker | FaceNet closed-set system, sticker dodging · Closed-set VGGFace2 subset | attack success rate: 0.01 -> 0.92 | Improved |
| Use grid-level face-mask rather than pixel-level sticker perturbations | Closed-set systems, zero-knowledge dodging · Closed-set VGGFace2 subset | attack success rate | Mixed |
| Attack open-set rather than closed-set recognition | Face recognition systems under zero-knowledge dodging · LFW and closed-set VGGFace2 subset | attack success rate | Improved |
| Use a universal face-mask perturbation rather than other universal perturbation types | Closed-set systems, N=20 · Closed-set VGGFace2 subset | attack success rate: 0.91 for FaceNet; 0.08 for ArcFace18 eyeglass-frame attack | Improved |
| Increase the number of inputs sharing one universal perturbation | Closed-set PGD, eyeglass-frame, and sticker attacks · Closed-set VGGFace2 subset | attack success rate: 0.94 -> 0.15 for VGGFace PGD; 0.98 -> 0.09 for VGGFace sticker | Hurt |
| Add momentum to the single-surrogate PGD transfer attack (ablation) | FaceNet closed-set system · Closed-set VGGFace2 subset, training-set knowledge | attack success rate: 0.42 -> 0.52 | Improved |
| Replace a single surrogate with an ensemble surrogate for PGD transfer (ablation) | FaceNet closed-set system · Closed-set VGGFace2 subset, training-set knowledge | attack success rate: 0.42 -> 0.73 | Improved |
| Add momentum to an ensemble-surrogate PGD transfer attack (ablation) | FaceNet closed-set system · Closed-set VGGFace2 subset, training-set knowledge | attack success rate: 0.73 -> 0.83 | Improved |
| Use ensemble and momentum transfer techniques for sticker attacks (ablation) | Closed-set face recognition systems · Closed-set VGGFace2 subset, training-set knowledge | attack success rate: 0.00 -> 0.01 on FaceNet | No clear effect |
| Apply DOA adversarial training against face-mask perturbations | Four of five closed-set systems · Closed-set systems, dodging | attack success rate | No clear effect |
| Apply DOA adversarial training to FaceNet against face-mask attacks | FaceNet closed-set system · Closed-set systems, dodging | attack success rate: 1.0 -> 0.24 | Improved |
| Generalize robustness from one perturbation type to another | DOA-retrained face recognition systems · Closed-set systems, physically realizable attacks | attack success rate | Mixed |
Closest papers by meaning
- Unravelling Robustness of Deep Learning Based Face Recognition Against Adversarial Attacks (AAAI 2018 · similarity 0.89)
- Towards Certifiably Robust Face Recognition (ECCV 2024 · similarity 0.88)
- Efficient Decision-Based Black-Box Adversarial Attacks on Face Recognition (CVPR 2019 · similarity 0.87)
- FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models (AAAI 2026 · similarity 0.85)
- Non-Adaptive Adversarial Face Generation (NeurIPS 2025 · similarity 0.84)
- Mixture-of-Attack-Experts with Class Regularization for Unified Physical-Digital Face Attack Detection (AAAI 2025 · similarity 0.84)
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.
