Toward Calibrated, Fair, and accurate Deepfake Detection

A new framework called Face-Fairness aims to close demographic performance gaps in deepfake detectors without requiring sensitive identity labels or retraining of the underlying models, according to a preprint posted to arXiv on 3 Jun 2026 [1]. The work, titled “Toward Calibrated, Fair, and accurate Deepfake Detection,” introduces a plug-and-play bias-mitigation system whose centerpiece is Face-Feature Tuning (FFT) [1][2]. FFT is described as a lightweight calibrator that remaps a detector’s output logits using frozen face embeddings, sidestepping the need for demographic attributes [2]. The authors note that existing fairness interventions typically demand demographic labels, full retraining, or trade away overall accuracy [2]. FFT is detector-agnostic, adds negligible runtime overhead, and requires no access to identity attributes [2]. The framework includes two complementary variants. FF-Max maximizes worst-group accuracy when demographic labels happen to be available, while FF-Discover achieves the same objective by inferring groups directly from the embedding space [2]. Across both in-domain and cross-dataset evaluations, the Face-Fairness family consistently reduced false-positive-rate and true-positive-rate gaps and lifted minimum group accuracy, often while preserving or improving overall accuracy [2]. Deepfake-detection fairness sits within a broader AI-alignment research agenda. Alignment work seeks to steer AI systems toward intended goals and ethical principles, and algorithmic fairness is recognized as a connected sub-discipline [3]. Biased or inaccurate training data has already been shown to make large language models less reliable, a parallel concern that underscores the importance of fairness interventions in generative-AI countermeasures [9]. The Face-Fairness preprint was posted on arXiv, an open-access repository that hosts e-prints after moderation but before formal peer review [7]. As of November 2024, the repository was receiving roughly 24,000 new articles per month [7].