Research map / Label-efficient and robust learning / Recognition and long-tail
Mitigating Biases in Blackbox Feature Extractors for Image Classification Tasks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Studies bias propagation from frozen blackbox pretrained feature extractors into downstream classifiers and proposes a bias-amplify-then-mitigate method using a clustering-based adaptive margin loss with randomized margins on an MLP adapter.
Problem
Debiasing downstream image classifiers built on frozen, blackbox pretrained feature encoders when only the classifier and a small adapter can be trained, and the bias attribute is unknown.
Approach
Amplify bias into an adapter via ERM with high weight decay, cluster the biased adapter features to estimate per-sample class-proportion-based margins, then train a debiased model with an ArcFace-style angular margin loss whose margins are Gaussian-randomized.
Key results
- On ResNet-18, the proposed margin-loss method reaches 80.29% worst-group accuracy on Waterbirds vs 67.57% for Contrastive Adapter and 38.90% for ERM.
- On CelebA (Blond Hair, ResNet-18) worst-group accuracy rises to 81.61% vs 27.20% ERM and 78.37% Co-Ada.
- On CMNIST-0.995 with ResNet-18, bias-conflicting accuracy is 72.56% vs 49.09% ERM and 65.48% Co-Ada.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply cluster-based adaptive margin loss | frozen ResNet-18 feature extractor with adapter · Waterbirds | worst-group accuracy: 67.57±1.29 -> 80.29±2.50 | Helped |
| apply cluster-based adaptive margin loss | frozen ResNet-18 feature extractor with adapter · CelebA Blond Hair | worst-group accuracy: 78.37±0.14 -> 81.61±1.02 | Helped |
| apply cluster-based adaptive margin loss | frozen ResNet-18 feature extractor with adapter · CMNIST-0.9 | average accuracy: 85.22±0.37 -> 87.33±0.35 | Helped |
| apply cluster-based adaptive margin loss | frozen ResNet-18 feature extractor with adapter · CMNIST-0.995 | bias-conflicting accuracy: 65.48±0.21 -> 72.56±0.88 | Helped |
| replace adaptive margins with a constant margin of 0.5 (ablation) | cluster-based margin method · Waterbirds | worst-group accuracy: 80.29 -> 55.76 | Hurt |
| remove Gaussian randomization of adaptive margins (ablation) | cluster-based margin method · Waterbirds | worst-group accuracy: 80.29 -> 79.28 | Hurt |
| remove Gaussian randomization of adaptive margins (ablation) | cluster-based margin method · CelebA Blond Hair | worst-group accuracy: 81.61 -> 78.51 | Hurt |
| cluster frozen encoder features instead of adapter features (ablation) | cluster-based margin method · Waterbirds | worst-group accuracy: 80.29 -> 74.92 | Hurt |
| use loss-weighted cross-entropy instead of cluster-based margin loss (ablation) | bias mitigation model · CelebA Blond Hair | worst-group accuracy: 81.61 -> 45.56 | Hurt |
| use cluster-weighted cross-entropy instead of cluster-based margin loss (ablation) | bias mitigation model · CelebA Blond Hair | worst-group accuracy: 81.61 -> 69.4 | Hurt |
Surprising findings
- Many existing debiasing methods (DebiAN, BPA, LfF, JTT, GEORGE) give inconsistent gains or none in the frozen-encoder setting, with LfF dropping 14% below ERM on CelebA worst-group accuracy.
- Clustering the frozen pretrained features instead of the adapter features gives only a modest drop in accuracy, so the frozen features can serve as a cheap proxy.
Limitations
- Relies on bias groups being categorical so that clustering can identify them.
- If the bias amplification or mitigation stage fails, the bias may be exacerbated.
- Method assumes the pretrained encoder's bias aligns with the downstream dataset bias; it fails when the encoder does not capture the downstream bias.
Experimental setup
Datasets: Waterbirds CelebA ColorMNIST UTKFace BAR
Benchmarks: Waterbirds CelebA ColorMNIST-0.9 ColorMNIST-0.995 UTKFace BAR
Metrics: worst-group accuracy average group accuracy bias-conflicting accuracy bias-aligned accuracy
Baselines: ERM DebiAN BPA LfF JTT GEORGE Contrastive Adapter (Co-Ada) DFR
Builds on: ArcFace Learning from Failure (LfF) Just Train Twice (JTT) Contrastive Adapter Deep Feature Reweighting (DFR)
Closest papers by meaning
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.87)
- Enhancing Intrinsic Features for Debiasing via Investigating Class-Discerning Common Attributes in Bias-Contrastive Pair (CVPR 2024 · similarity 0.85)
- Discover and Mitigate Multiple Biased Subgroups in Image Classifiers (CVPR 2024 · similarity 0.84)
- A Decade's Battle on Dataset Bias: Are We There Yet? (ICLR 2025 · similarity 0.84)
- Common Sense Bias Modeling for Classification Tasks (AAAI 2025 · similarity 0.84)
- Classes Are Not Equal: An Empirical Study on Image Recognition Fairness (CVPR 2024 · 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.