Content-Invariant Spatio-Temporal Neural Framework for Forgery Detection in Image Sequences

Content-Invariant Spatio-Temporal Neural Framework for Forgery Detection in Image Sequences

Paper in proceedings (conference paper)

The increasing prevalence of deepfake videos underscores the need for effective and reliable detection methods. In this study, we propose a hybrid deepfake detection framework that integrates a static image forgery detector with a recurrent neural network (RNN) to exploit both spatial and temporal fea- tures. Specifically, we utilize an existing frame-level detector that identifies common forgery artifacts within individual frames. This is followed by a Long Short-Term Memory (LSTM) network that models temporal dependencies across frames, enabling detection of inconsistencies that are overlooked in frame-by-frame anal- ysis. Experimental results demonstrate that temporal modeling significantly improves accuracy over frame-level baselines. Our contributions are twofold: (i) we provide empirical evidence that deepfake videos exhibit detectable temporal signatures, and (ii) we construct a compact, real-world evaluation set of deepfake videos. Notably, detection performance on this dataset is lower than on standard benchmarks, suggesting a domain gap between commonly used training data and real-world deepfakes.

The increasing prevalence of deepfake videos underscores the need for effective and reliable detection methods. In this study, we propose a hybrid deepfake detection framework that integrates a static image forgery detector with a recurrent neural network (RNN) to exploit both spatial and temporal fea- tures. Specifically, we utilize an existing frame-level detector that identifies common forgery artifacts within individual frames. This is followed by a Long Short-Term Memory (LSTM) network that models temporal dependencies across frames, enabling detection of inconsistencies that are overlooked in frame-by-frame anal- ysis. Experimental results demonstrate that temporal modeling significantly improves accuracy over frame-level baselines. Our contributions are twofold: (i) we provide empirical evidence that deepfake videos exhibit detectable temporal signatures, and (ii) we construct a compact, real-world evaluation set of deepfake videos. Notably, detection performance on this dataset is lower than on standard benchmarks, suggesting a domain gap between commonly used training data and real-world deepfakes.

deepfake detection, computer vision, spatio-temporal features, recurrent neural networks, image manipulation.

deepfake detection, computer vision, spatio-temporal features, recurrent neural networks, image manipulation.

BUZOVSKÝ, V.; PŘINOSIL, J.; ŘÍHA, K.; SMÉKAL, Z.

03.12.2025

979-8-3315-7675-2

International Conference on Ultra Modern Telecommunications and Workshops

International Congress on Ultra Modern Telecommunications and Workshops

United States of America

240

245

5