Signal, Image and Video Processing Signal Processing

Non-linear latent space expansion for generalizable image forensics and deepfake detection

Sahotra, Nikhil, Raikwar, Suresh, Glencross, Mashhuda, Singh, Priyanka
Springer Nature · 2026-11-16 · DOI 10.1007/s11760-026-05743-2

Abstract

The rise in realistic deep-fake video poses serious challenges to the trustworthiness of digital media. The introduction of generative AI techniques and its ongoing evolution create a challenge for deep fake detection. Existing detectors struggle to identify new types of synthesized content as video generation methods improve. To tackle this problem, the proposed method uses extends the latent space non-lineraly. This improves the generalization of deepfake detection by mapping complex forgery spaces and expanding the latent space to represent diverse unseen manipulation patterns. The proposed method is four folds: (1) uses augmentation techniques to handle a variety of fake scenarios, (2) improves understanding of real features by combining EfficientNet-B4 and ResNet100 into a single strong encoder, (3) uses SSIM-based distillation loss to better preserve the structural and perceptual quality of features, and (4) incorporates explainable AI to indicate the reasons of an image be fake/real. The proposed method has been evaluated using benchmark datasets and compared with state-of-the-art methods.

Read at publisher →