ACL Findings2023
Shielded Representations: Protecting Sensitive Attributes Through Iterative Gradient-Based Projection
Abstract
Natural language processing models tend to learn and encode social biases present in the data. One popular approach for addressing such biases is to eliminate encoded information from the model's representations. However, current methods are restricted to removing only linearly encoded information. In this work, we propose Iterative Gradient-Based Projection (IGBP), a novel method for removing non-linear encoded concepts from neural representations. Our method consists of iteratively training neural classifiers to predict a particular attribute we seek to eliminate, followed by a projection of the representation on a hypersurface, such that the classifiers become oblivious to the target attribute. We evaluate the effectiveness of our method on the task of removing gender and race information as sensitive attributes. Our results demonstrate that IGBP is effective in mitigating bias through intrinsic and extrinsic evaluations, with minimal impact on downstream task accuracy. 1
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
8 top-tier papers have cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- Probing Classifiers are Unreliable for Concept Removal and Detection cites Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection
- Precise In-Parameter Concept Erasure in Large Language Models cites Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection
- Spectral Editing of Activations for Large Language Model Alignment cites Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection
- Spectral Attention Steering for Prompt Highlighting cites Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection
- Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning cites Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection
- Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuning cites Double-Hard Debias: Tailoring Word Embeddings for Gender Bias Mitigation
Cited by top-tier papers8
- Spectral Editing of Activations for Large Language Model AlignmentYifu Qiu, Zheng Zhao, Yftah Ziser, Anna Korhonen et al.NeurIPS 2024 · 66 citations
- Steering Out-of-Distribution Generalization with Concept Ablation Fine-TuningHelena Casademunt, Caden Juang, Adam Karvonen, Samuel Marks et al.ICML 2026 · 32 citations
- Precise In-Parameter Concept Erasure in Large Language ModelsYoav Gur-Arieh, Clara Suslik, Yihuai Hong, Fazl Barez et al.EMNLP 2025 · 10 citations
- Spectral Attention Steering for Prompt HighlightingWeixian Waylon Li, Yuchen Niu, Yongxin Yang, Keshuang Li et al.ICLR 2026 · 6 citations
- FALCON: Fine-grained Activation Manipulation by Contrastive Orthogonal Unalignment for Large Language ModelJinwei Hu, Zhenglin Huang, Xiangyu Yin, Wenjie Ruan et al.NeurIPS 2025 · 3 citations
Builds on6
- Balanced Datasets Are Not Enough: Estimating and Mitigating Gender Bias in Deep Image RepresentationsTianlu Wang, Jieyu Zhao, Mark Yatskar, Kai-Wei Chang et al.ICCV 2019 · 469 citations
- Probing Classifiers are Unreliable for Concept Removal and DetectionAbhinav Kumar, Chenhao Tan, Amit SharmaNeurIPS 2022 · 46 citations
- Double-Hard Debias: Tailoring Word Embeddings for Gender Bias MitigationTianlu Wang, Xi Victoria Lin, Nazneen Fatema Rajani, Bryan McCann et al.ACL 2020 · 42 citations
- Information-Theoretic Probing with Minimum Description LengthElena Voita, Ivan TitovEMNLP 2020 · 34 citations
- Null It Out: Guarding Protected Attributes by Iterative Nullspace ProjectionShauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton et al.ACL 2020 · 25 citations
Related papers
- Linear Adversarial Concept ErasureShauli Ravfogel, Michael Twiton, Yoav Goldberg, Ryan CotterellICML 2022 · 89 citations
- Preserving Task-Relevant Information Under Linear Concept RemovalFloris Holstege, Shauli Ravfogel, Bram WoutersNeurIPS 2025 · 4 citations
- GRADIEND: Feature Learning within Neural Networks Exemplified through BiasesJonathan Drechsel, Steffen HerboldICLR 2026 · 2 citations
- Adversarial Concept Erasure in Kernel SpaceShauli Ravfogel, Francisco Vargas, Yoav Goldberg, Ryan CotterellEMNLP 2022 · 11 citations
- Unsupervised Concept Vector Extraction for Bias Control in LLMsHannah Cyberey, Yangfeng Ji, David EvansEMNLP 2025 · 4 citations
