定位CLIP模型中性别与年龄偏见的注意力头，揭示其在视觉编码器中的具体位置。
Locating Demographic Bias at the Attention-Head Level in CLIP's Vision Encoder
- 通过注意力头级分解和概念激活向量，定位模型内部偏见来源。
- 删除特定注意力头可降低性别偏见（Cramer's V从0.381降至0.362），且准确率微升0.42%。
- 性别偏见集中在末层少数注意力头，而年龄偏见分布更分散，难以通过单一头消除。
主流公平性审计仅能判断基础模型存在偏见，却无法定位偏见在模型内部的具体位置。本文提出一种机制性公平审计方法，结合投影残差流分解、零样本概念激活向量与增强文本跨度分析，实现对视觉变换器中个体注意力头层面的性别与年龄偏见定位。以FACET基准的42个职业类别为例，对CLIP ViT-L-14编码器进行评估。在性别偏见方面，识别出四个末层注意力头，其移除使全局偏见下降（Cramer's V：0.381 → 0.362），准确率小幅提升0.42%；层匹配随机对照验证了该效应具有针对性。其中单个末层头主导了最刻板职业类别的偏见修正，类别级分析显示预测结果更趋正确。在年龄偏见方面，虽识别出候选注意力头，但移除效果较弱且不一致，表明年龄偏见在模型中编码更弥散。结果初步证实，对判别性视觉编码器而言，注意力头级偏见定位是可行的，且不同受保护属性的可定位性存在差异。
原文摘要 · Abstract (English)
Standard fairness audits of foundation models quantify that a model is biased, but not where inside the network the bias resides. We propose a mechanistic fairness audit that combines projected residual-stream decomposition, zero-shot Concept Activation Vectors, and bias-augmented TextSpan analysis to locate demographic bias at the level of individual attention heads in vision transformers. As a feasibility case study, we apply this pipeline to the CLIP ViT-L-14 encoder on 42 profession classes of the FACET benchmark, auditing both gender and age bias. For gender, the pipeline identifies four terminal-layer heads whose ablation reduces global bias (Cramer's V: 0.381 -> 0.362) while marginally improving accuracy (+0.42%); a layer-matched random control confirms that this effect is specific to the identified heads. A single head in the final layer contributes to the majority of the reduction in the most stereotyped classes, and class-level analysis shows that corrected predictions shift toward the correct occupation. For age, the same pipeline identifies candidate heads, but ablation produces weaker and less consistent effects, suggesting that age bias is encoded more diffusely than gender bias in this model. These results provide preliminary evidence that head-level bias localisation is feasible for discriminative vision encoders and that the degree of localisability may vary across protected attributes. keywords: Bias . CLIP . Mechanistic Interpretability . Vision Transformer . Fairness
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。