From Deception to Perception: The Surprising Benefits of Deepfakes for Detecting, Measuring, and Mitigating Bias

View on arXiv ← Back to list

Authors: Yizhi Liu, Balaji Padmanabhan, Siva Viswanathan

Published: 2025-02-16 16:55:28+00:00

AI Summary

This research uses deepfake technology to generate controlled facial images for correspondence studies, extending their scope beyond textual manipulations to assess biases in areas like pain assessment. The results demonstrate deepfakes' effectiveness in measuring and potentially correcting biases related to race and age in pain assessment.

Abstract

While deepfake technologies have predominantly been criticized for potential misuse, our study demonstrates their significant potential as tools for detecting, measuring, and mitigating biases in key societal domains. By employing deepfake technology to generate controlled facial images, we extend the scope of traditional correspondence studies beyond mere textual manipulations. This enhancement is crucial in scenarios such as pain assessments, where subjective biases triggered by sensitive features in facial images can profoundly affect outcomes. Our results reveal that deepfakes not only maintain the effectiveness of correspondence studies but also introduce groundbreaking advancements in bias measurement and correction techniques. This study emphasizes the constructive role of deepfake technologies as essential tools for advancing societal equity and fairness.

Key findings

Significant racial and age biases were found in pain assessments, with higher pain scores consistently assigned to white and older individuals. Averaging labels from original and race-manipulated images in AI model training significantly improved individual fairness, suggesting a novel approach to bias correction.

Approach

The researchers used GAN-based deepfake techniques to manipulate facial images (race, age, or both) while preserving pain expression. They then conducted experiments on crowdsourcing platforms (AMT and Credamo) to assess pain levels by human assessors and used these results to train and evaluate AI models for bias detection and correction.

Datasets

100 facial images of painful expressions from a digital pain management company, categorized by race (Black, White), age (young, senior), and gender. Deepfakes were used to generate additional images, resulting in a total of 400 images.

Model(s)

StyleGAN, StyleFeatureEditor, E4e, StyleCLIP, InterfaceGAN, ResNet50, VGGFace, BlazeFace, a state-of-the-art Vision Transformer (ViT) model for deepfake detection.

Author countries

USA

← Previous