Physics-Informed Deepfake Detection in Facial Images Using Landmark Geometry and Anatomy-Aware Hybrid Classification

DOI

10.2991/978-94-6239-707-1_21How to use a DOI?

Keywords

deepfake detection; facial landmarks; physics-informed learning; cross-dataset robustness; interpretable deep learning

Abstract

Deepfake detectors trained on a single dataset often fail under unseen manipulations because they rely on dataset-specific artifacts rather than facial characteristics. We propose a physics-informed deepfake detector that grounds decision-making in anatomical invariants of real faces, such as bilateral symmetry, smooth contours, and characteristic proportions, while deepfakes often break these patterns due to synthesis artifacts. A multi-stream CNN-GNN module extracts robust landmarks, and five anatomy-aware descriptor families measure physics violations in a hybrid geometry-appearance classifier. Evaluated across three cross-dataset protocols on FFHQ, CelebA, and FaceForensics++, the model achieves 95.8% accuracy on FaceForensics++ and 92.3% mean cross-dataset accuracy averaged over FF++→FFHQ, FF++→CelebA, and cross-manipulation protocols, outperforming a CNN baseline by 11.2 percentage points and reducing cross-dataset variance by about 78%. The proposed framework shows greatly enhanced robustness to manipulations including Face2Face, FaceSwap, reenactment, and diffusion, with an accuracy standard deviation of 2.1% compared to 9.4% for CNN baselines. The proposed system reduces facial landmark symmetry error by 18.7% over MediaPipe through physics-based correction of geometric inaccuracies.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Soumitra Ghosh
AU  - Sudhir Ranjan Pattanaik
PY  - 2026
DA  - 2026/06/18
TI  - Physics-Informed Deepfake Detection in Facial Images Using Landmark Geometry and Anatomy-Aware Hybrid Classification
BT  - Proceedings of the International Conference on Recent Advances in Intelligent and Sustainable Technologies (RAIST 2026)
PB  - Atlantis Press
SP  - 243
EP  - 253
SN  - 2589-4919
UR  - https://doi.org/10.2991/978-94-6239-707-1_21
DO  - 10.2991/978-94-6239-707-1_21
ID  - Ghosh2026
ER  -