FairReL: Deepfake Detection using Fairness-Aware Representation Learning

Authors: Xiaoman Lu, Jiaqi Li, Shuntian Zheng, Huiping Chen, Yu Guan

Published: 2026-08-28 18:34:15+00:00

Comment: 26 pages, 7 figures. Accepted at BMVC 2026. Code: https://github.com/xiaoman89/FairReL

AI Summary

This paper introduces FairReL, a fairness-aware representation-learning framework for deepfake detection that addresses uneven error distribution across demographic subgroups. FairReL targets two subgroup-sensitive components: multi-scale spatial features and fine-tuning-induced residual features, using dedicated demographic supervision through Group-Conditional Wavelet Decorrelation (GCWD) and Subspace-Localised Mean Alignment (SLMA). It aims to improve unseen-dataset AUC while significantly reducing subgroup False Positive Rate (FPR) disparity compared to state-of-the-art fairness-aware detectors.

Abstract

Although recent deepfake detectors achieve high overall accuracy, their errors remain unevenly distributed across demographic subgroups, with real faces from certain groups more often misclassified as fake. Existing fairness-aware detectors typically regularise the entire feature representation, without identifying or controlling the specific components that drive unfair predictions. Such coarse intervention can over-suppress useful forgery cues while leaving demographic structure in component-specific subspaces. To address this, we identify two subgroup-sensitive components: multi-scale spatial features, which encode local facial and forgery patterns, and fine-tuning-induced residual features, which adapt the backbone to the unfair training distribution. We propose FairReL, a fairness-aware representation-learning framework that targets both components with dedicated demographic supervision. FairReL uses an SVD-decomposed foundation-model backbone to isolate the fine-tuning-induced residual representation, and introduces two complementary losses. Group-Conditional Wavelet Decorrelation (GCWD) suppresses subgroup-imbalanced structure across spatial wavelet sub-bands, while Subspace-Localised Mean Alignment (SLMA) aligns subgroup means within each real/fake class in the residual representation. Experiments on FF++, Celeb-DF, DFD and DFDC show that, against the state-of-the-art fairness-aware detector, FairReL improves unseen-dataset AUC by 3.9% while reducing subgroup FPR disparity by 10.2%. Code is available at https://github.com/xiaoman89/FairReL .

Key findings

FairReL significantly improves unseen-dataset AUC by 3.9% and reduces subgroup FPR disparity by 10.2% compared to state-of-the-art fairness-aware detectors. It achieves the lowest False Positive Rate (FFPR) and Max Equalized Odds (FMEO) on all three unseen datasets while maintaining competitive AUC. Mechanistic analyses confirm that GCWD effectively reduces low-frequency subgroup discrepancy, and SLMA suppresses adaptation-induced subgroup shifts concentrated in the residual pathway.

Approach

FairReL identifies and targets two specific components responsible for unfair predictions: multi-scale spatial features and fine-tuning-induced residual features. It uses an SVD-decomposed foundation-model backbone to isolate the residual representation. GCWD suppresses subgroup-imbalanced structure in spatial wavelet sub-bands, while SLMA aligns subgroup means within each real/fake class in the residual representation, reducing adaptation-induced demographic shifts.

Datasets

FaceForensics++ (FF++), Celeb-DF, DFD, DFDC

Model(s)

CLIP ViT-L/14 (with SVD decomposition), Xception

Author countries

United Kingdom

← Previous