Intersectional Invariant Learning for Fair Deepfake Detection

Authors: Uzoamaka Ezeakunne, Taiwo Olabamipe, Deevesh Jain, Xiuwen Liu

Published: 2026-10-27 19:22:19+00:00

Comment: 4 pages including appendices, 7 figures, 8 tables

AI Summary

This paper introduces Intersectional Invariant Learning (IIL), a framework designed to mitigate performance disparities in deepfake detection across demographic groups by addressing underlying feature-level bias. IIL achieves this by iteratively identifying challenging demographic partitions and learning features that are consistent across them, incorporating a false-positive-rate fairness objective and group-balanced sampling. The framework demonstrates Pareto-optimal performance on a large-scale fairness benchmark, simultaneously achieving high accuracy and significantly reducing intersectional fairness disparities.

Abstract

Deepfake detection models achieve high overall accuracy but exhibit large performance disparities across demographic groups, causing some groups to experience substantially higher false-positive rates than others. Existing fairness-aware methods address this by constraining model outputs to equalize group predictions, but do not tackle the underlying feature-level bias. As a result, models may still rely on demographic characteristics, such as skin texture, as indirect cues for detecting manipulated content. To address this problem, we propose Intersectional Invariant Learning (IIL), a framework that improves feature learning rather than only adjusting model outputs. IIL alternates between identifying the most challenging demographic partitions and learning features that remain consistent across those partitions. The framework also includes a false-positive-rate fairness objective and group-balanced sampling across six gender x skin-tone intersectional groups. We evaluate IIL on a large-scale fairness benchmark containing over one million images, 10 baseline detectors, and 21 fairness and utility metrics. Results show that IIL achieves Pareto-optimal performance, simultaneously attaining an AUC above 99%, a false-positive rate under 2.5%, and the lowest intersectional fairness disparity (FMEO) among all compared methods. IIL outperforms prior fairness-aware methods on the large majority of the 21 evaluation metrics, while reducing both false-positive rates and intersectional fairness disparities by several-fold compared with the strongest competing method.


Key findings
IIL achieved Pareto-optimal performance, with an AUC above 99% and a false-positive rate under 2.5%, while exhibiting the lowest intersectional fairness disparity (FMEO) among all compared methods. It significantly reduced both false-positive rates and intersectional fairness disparities by several-fold compared to the strongest competing methods, demonstrating improved generalization across various datasets.
Approach
IIL alternates between identifying the most challenging demographic partitions using Risk Extrapolation (REx) and learning features that remain consistent across those partitions using Invariant Risk Minimization (IRMv1). The framework also includes a false-positive-rate fairness objective (LFPR) and group-balanced sampling across six gender x skin-tone intersectional groups to directly target false-positive disparity and prediction-rate parity.
Datasets
AI-Face (including constituent sources: Celeb-DF, DFD, DFDC, and FF++)
Model(s)
Xception
Author countries
UNKNOWN