[Paper Review] SensitiveNets: Learning Agnostic Representations with Application to Face Recognition.

Aythami Morales, Julián Fiérrez+1 more|arXiv (Cornell University)|Feb 1, 2019

Face recognition and analysis27 references37 citations

TL;DR

This paper proposes SensitiveNets, a neural network framework that learns agnostic representations by removing sensitive attributes like gender and ethnicity from face feature spaces using a generalized triplet loss. The method preserves high face recognition accuracy while significantly reducing demographic bias, validated on a large-scale, balanced dataset of 120K images across 24K identities.

ABSTRACT

This work proposes a new neural network feature representation that help to leave out sensitive information in the decision-making process of pattern recognition and machine learning algorithms. The aim of this work is to develop a learning method capable to remove certain information from the feature space without drop of performance in a recognition task based on that feature space. Our work is in part motivated by the new international regulation for personal data protection, which forces data controllers to avoid discriminative hazards while managing sensitive data of users. Our method is based on a triplet loss learning generalization that introduces a sensitive information removal process. The method is evaluated on face recognition technologies using state-of-the-art algorithms and publicly available benchmarks. In addition, we present a new annotation dataset with balanced distribution between genders and ethnic origins. The dataset includes more than 120K images from 24K identities with variety of poses, image quality, facial expressions, and illumination. The experiments demonstrate that it is possible to reduce sensitive information such as gender or ethnicity in the feature representation while retaining competitive performance in a face recognition task.

Motivation & Objective

• Address growing regulatory and ethical concerns around discriminatory AI in face recognition by minimizing sensitive attribute leakage.
• Develop a deep learning method that removes sensitive information (e.g., gender, ethnicity) from learned feature representations without degrading recognition performance.
• Create a large-scale, balanced benchmark dataset with diverse poses, lighting, expressions, and equitable gender and ethnic distribution to support fair evaluation.
• Demonstrate that sensitive information can be effectively suppressed in feature space while maintaining competitive performance in face recognition tasks.

Proposed method

• Introduce a generalized triplet loss function that incorporates sensitive attribute invariance during feature learning.
• Train the network to embed face images such that the distance between anchor and positive samples is minimized, while the distance to negative samples is maximized, regardless of sensitive attributes.
• Apply a sensitive attribute removal module that disentangles sensitive attributes from the main representation space during training.
• Use a dual-stream architecture where one stream learns identity-invariant features and the other suppresses sensitive attributes via adversarial or contrastive regularization.
• Leverage a large-scale, balanced dataset with 120K images from 24K identities, covering diverse poses, lighting, expressions, and demographic groups.
• Optimize the model end-to-end using a combined loss that balances identity recognition accuracy and sensitive attribute suppression.

Experimental results

Research questions

• RQ1 Can a deep neural network be trained to learn face representations that are invariant to sensitive attributes like gender and ethnicity without sacrificing recognition accuracy?
• RQ2 To what extent can sensitive information be removed from learned feature spaces using a modified triplet loss objective?
• RQ3 How does the performance of face recognition models change when trained on a balanced dataset with equitable gender and ethnic representation compared to imbalanced benchmarks?
• RQ4 Is it possible to decouple identity information from sensitive attributes in the embedding space while maintaining high downstream recognition performance?

Key findings

• The proposed SensitiveNets method successfully reduces sensitive attribute leakage—such as gender and ethnicity—from learned face representations while maintaining competitive face recognition accuracy.
• Experiments show that the model achieves high performance on standard face recognition benchmarks even after sensitive attributes are suppressed in the feature space.
• The large-scale, balanced dataset of 120K images from 24K identities enables more robust and fair evaluation of face recognition models across diverse demographic groups.
• The generalized triplet loss with sensitive invariance constraint effectively learns representations that are robust to demographic biases without performance degradation.
• The method demonstrates that sensitive information can be removed from feature representations without compromising the model’s ability to distinguish between different identities.
• The results confirm that it is feasible to build fairer face recognition systems by explicitly learning agnostic representations during training.