Synthetic Counterfactual Faces

TL;DR

This work tackles the challenge of evaluating fairness and robustness of face understanding systems under semantic distribution shifts by generating targeted counterfactual faces. It introduces an end-to-end diffusion-based pipeline that creates source faces and applies semantic edits via a latent manipulator, with filtering guided by distortion and attribute detectors. The authors generate about 15k counterfactual faces across 8 demographic groups and 19 attributes, validate the data with human studies, and demonstrate a practical vulnerability in a leading commercial model (Instagram Android Image Understanding). The study provides a controllable, validated dataset and methodology for targeted bias analysis and robustness evaluation in real-world vision systems.

Abstract

Computer vision systems have been deployed in various applications involving biometrics like human faces. These systems can identify social media users, search for missing persons, and verify identity of individuals. While computer vision models are often evaluated for accuracy on available benchmarks, more annotated data is necessary to learn about their robustness and fairness against semantic distributional shifts in input data, especially in face data. Among annotated data, counterfactual examples grant strong explainability characteristics. Because collecting natural face data is prohibitively expensive, we put forth a generative AI-based framework to construct targeted, counterfactual, high-quality synthetic face data. Our synthetic data pipeline has many use cases, including face recognition systems sensitivity evaluations and image understanding system probes. The pipeline is validated with multiple user studies. We showcase the efficacy of our face generation pipeline on a leading commercial vision model. We identify facial attributes that cause vision systems to fail.

Figures (8)

• Figure 1: Our counterfactual image generation pipeline. Only images that are strictly counterfactuals pass through the verification portion.
• Figure 2: Generating transformed images with Instruction-guided editing methods 'InstructPix2Pix', 'MagicBrush', and 'HQ-Edit': All edits were with a text guidance of 7.5, and image guidance of 1.5. The first column contains the original face(source face)
• Figure 3: Attribute Change Transititon Matrix for the Candiate Filtering Step: This matrix contains our requirements for what should happen to all the attributes on a transformed face obtained from a source face while inducing an attribute. The columns indicate all the attributes and each row is the attribute being applied. '1' indicates attribute should be present in the transformed face. '0' indicates attribute should not be present in the transformed face. '-1' indicates that attribute should be present in the transformed face only if it is there in the source face. '-2' indicates attribute is not considered for the filtering.
• Figure 4: Subset of distorted faces used in training distortion detector
• Figure 5: Example of images labeled as distorted in the Distortion Survey