Back to search

University of North Texas

DF-OOD: Real-Only Deepfake Detection Via Confidence Dynamics under Perturbations

Abstract

dc:description

Existing deepfake detection methods predomi- nantly rely on supervised learning with labeled real and manipulated data, limiting their ability to generalize to un- seen manipulation techniques. In this work, we reformulate deepfake detection as an out-of-distribution (OOD) problem and propose DF-OOD, a one-class framework that learns the distribution of authentic facial images using real-only training. At inference, manipulated samples are identified as deviations from the learned real-face manifold. To enhance separability between in-distribution and anomalous samples, we introduce a confidence-driven perturbation mechanism that leverages temperature scaling and gradient-based input pertur- bations to amplify differences in model response. We observe that authentic images exhibit stable and consistently increasing confidence under perturbations, whereas manipulated images produce weaker and less stable responses due to their deviation from the learned distribution. We evaluate DF-OOD across several diverse datasets spanning identity swap, reenactment, GAN-based synthesis, and diffusion-based generation, including FaceForensics++, CelebDF, DeepFakeFace, DF40, StyleGAN, StarGAN, and PGGAN. Experimental results demonstrate strong cross-manipulation generalization without requiring any manipulated data during training, achieving AUROC up to 0.99 on diffusion-based deepfakes and above 0.94 across all Face- Forensics++ manipulation types, while remaining competitive with fully supervised baselines.

Degree

thesis:*
Grantor dc:publisher
University of North Texas
Year dc:date
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Patel, Miliben Alpeshkumar 2002-
Contributors dc:contributor
  • Tunc, Cihan
  • Rattani, Ajita
  • Morozov, Kirill

Subjects

dc:subject × 3

Rights

dc:rights
Statement dc:rights
  • Public
  • Patel, Miliben Alpeshkumar 2002-
  • Copyright
  • Copyright is held by the author, unless otherwise noted. All rights Reserved.
Language dc:language
English

Identifiers

dc:identifier.*
Identifier
local-cont-no: submission_4924
ark: ark:/67531/metadc2652434

Chain of custody

source
Harvested from
University of North Texas
Base URL
digital.library.unt.edu/oai/
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Patel, Miliben Alpeshkumar 2002-. DF-OOD: Real-Only Deepfake Detection Via Confidence Dynamics under Perturbations. University of North Texas, 2026. https://digital.library.unt.edu/ark:/67531/metadc2652434/