[论文解读] Deep Learning for Face Recognition: Pride or Prejudiced?

Shruti Nagpal, Maneet Singh+2 more|arXiv (Cornell University)|Apr 2, 2019

Face recognition and analysis参考文献 33被引用 39

一句话总结

本文分析了最先进的深度学习人脸识别模型是否编码自身种族偏见和自身年龄偏见，并考察偏见在四个网络和36个实验中的表现位置与方式。

ABSTRACT

Do very high accuracies of deep networks suggest pride of effective AI or are deep networks prejudiced? Do they suffer from in-group biases (own-race-bias and own-age-bias), and mimic the human behavior? Is in-group specific information being encoded sub-consciously by the deep networks? This research attempts to answer these questions and presents an in-depth analysis of `bias' in deep learning based face recognition systems. This is the first work which decodes if and where bias is encoded for face recognition. Taking cues from cognitive studies, we inspect if deep networks are also affected by social in- and out-group effect. Networks are analyzed for own-race and own-age bias, both of which have been well established in human beings. The sub-conscious behavior of face recognition models is examined to understand if they encode race or age specific features for face recognition. Analysis is performed based on 36 experiments conducted on multiple datasets. Four deep learning networks either trained from scratch or pre-trained on over 10M images are used. Variations across class activation maps and feature visualizations provide novel insights into the functioning of deep learning systems, suggesting behavior similar to humans. It is our belief that a better understanding of state-of-the-art deep learning networks would enable researchers to address the given challenge of bias in AI, and develop fairer systems.

研究动机与目标

• 评估深度学习人脸识别是否编码种族特异信息（own-race bias）和年龄特异信息（own-age bias）。
• 识别不同子群体网络依赖的面部区域以及训练数据分布如何影响偏见。
• 研究预训练和微调如何影响偏见与感兴趣区域模式。
• 提供见解以引导去偏见化和在面部识别领域实现更公平的AI。

提出的方法

• 分析四个深度学习人脸识别网络（LightCNN-9, LightCNN-29, ResNet50, SENet50）在从零开始训练或在大规模数据集上预训练后进行训练。
• 使用类激活映射（CAMs）和特征可视化来识别不同子群体间的判别性面部区域。
• 在 Race-A 与 Race-B 的对照组以及 0–14、15–32、33+ 的年龄组之间进行36个受控实验，训练集/测试集分离独立。
• 在基于种族和年龄的子群体上比较验证精度（FAR 1% 时的 Genuine Acceptance Rate）。
• 考察从零开始训练、在大数据集上的预训练、跨子群测试，以及对感兴趣区域的微调对偏见和区域焦点的影响。

实验结果

研究问题

• RQ1 深度学习人脸识别是否编码种族特异信息，从而导致自身种族偏见（own-race bias）？
• RQ2 深度学习是否编码年龄特异信息，从而导致自身年龄偏见（own-age bias）？
• RQ3 网络在识别不同种族或年龄组个体时依赖哪些面部区域？
• RQ4 预训练和微调如何影响子群体之间的偏见和区域关注点？

主要发现

• 在单一种族上训练的网络对其他种族的泛化能力较差，表明学习特征中存在自身种族偏见。
• 预训练网络在种族上表现出偏 biased，在跨离散数据集时对 Race-A 的准确率高于 Race-B。
• CAM 分析中出现种族特异的感兴趣区域，支持模型潜意识地编码了种族信息。
• 存在自身年龄偏见的证据，各网络对 youngest age group (0–14) 的识别表现下降；预训练有帮助但未完全缓解。
• 微调会改变感兴趣区域，有时聚焦于子群体特有的面部区域，可能降低泛化性。
• 大规模预训练提升总体准确性，但未消除子群体偏见，数据分布被确认为关键因素。