Research map / Adversarial robustness and security / Certified and empirical robustness
BiasAdv: Bias-Adversarial Augmentation for Model Debiasing
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 |
|---|---|---|---|
| Add BiasAdv to LfF | LfF · CIFAR-10C, p=0.5%, 2%, and 5% | AVERAGE accuracy: 28.81±0.44 -> 36.78±0.20; 40.66±0.70 -> 48.36±0.59; 50.72±1.31 -> 57.78±0.33 | Helped |
| Add BiasAdv to LfF | LfF · BFFHQ | CONFLICTING accuracy: 62.97±3.22 -> 72.40±1.34 | Helped |
| Add BiasAdv to LfF | LfF · BAR, p=1% and p=5% | CONFLICTING accuracy: 57.71±3.12 -> 63.20±2.64; 67.48±0.46 -> 72.62±0.11 | Helped |
| Add BiasAdv to LfF | LfF · MetaShift, p=1%, 6%, and 12% | WORST-GROUP accuracy: 52.78±2.04 -> 53.47±2.91; 57.17±1.99 -> 58.10±1.43; 62.73±2.29 -> 64.35±1.45 | Helped |
| Add BiasAdv to ERM | ERM · MetaShift, p=6% | AVERAGE accuracy: 79.69±0.86 -> 78.70±1.71 | No clear effect |
| Replace BiasAdv images with random-noise augmentation (ablation) | ERM and LfF · BFFHQ | CONFLICTING accuracy: ERM: 57.73±0.19 -> 55.00±0.55; LfF: 72.40±1.34 -> 64.27±1.27 | Hurt |
| Attack the debiased model instead of the auxiliary model (ablation) | ERM and LfF · BFFHQ | CONFLICTING accuracy: ERM: 57.73±0.19 -> 50.68±1.03; LfF: 72.40±1.34 -> 60.80±1.05 | Hurt |
| Remove the prediction-preservation regularization term (ablation) | BiasAdv with ERM and LfF · BFFHQ | CONFLICTING accuracy: ERM: 57.73±0.19 -> 56.16±0.74; LfF: 72.40±1.34 -> 66.13±2.17 | Hurt |
Closest papers by meaning
- MABR: Multilayer Adversarial Bias Removal Without Prior Bias Knowledge (AAAI 2025 · similarity 0.90)
- A Simple Remedy for Dataset Bias via Self-Influence: A Mislabeled Sample Perspective (NeurIPS 2024 · similarity 0.89)
- Diffusing DeBias: Synthetic Bias Amplification for Model Debiasing (NeurIPS 2025 · similarity 0.89)
- DeNetDM: Debiasing by Network Depth Modulation (NeurIPS 2024 · similarity 0.88)
- Self-supervised Debiasing Using Low Rank Regularization (CVPR 2024 · similarity 0.88)
- Enhancing Intrinsic Features for Debiasing via Investigating Class-Discerning Common Attributes in Bias-Contrastive Pair (CVPR 2024 · 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.