TL;DR
This work proposes a spatial-temporal model with two key components: one targeting artifacts within individual frames and the other analyzing inconsistencies across consecutive frames, both leveraging a bidirectional long short-term memory (Bi-LSTM) mechanism.
This work proposes a parameter-efficient, video-based deepfake detection framework that leverages a frozen foundation model encoder coupled with a lightweight spatio-temporal decoder, complemented by a Bidirectional Spatio-Temporal decoder that models local temporal transitions and bidirectional temporal dependencies across sampled frames, enabling robust temporal reasoning within each video clip.
BlinkNet, an explainable deepfake-detection framework that examines the spatial appearance and temporal kinematics of eye blinks, indicates that ocular dynamics can complement spatial evidence while improving efficiency and interpretability.
Deepfake videos generated using modern deep learning techniques pose significant threats to digital media authenticity and public trust. These manipulated videos often appear highly realistic, making manual verification difficult. This paper proposes an efficient spatiotemporal deepfake detection framework that combines Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and a temporal attention mechanism. MobileNetV2 is utilized as a CNN model for spatial feature extraction in individual video frames. The extracted features are then subjected to the LSTM network for detection of temporal inconsistencies in the video frames. Furthermore, the use of the temporal attention mechanism is proposed for the identification of video frames where the level of manipulation is higher. The experimental results show the efficiency and accuracy of the proposed method in video forgery detection. The proposed method attains 96.1% detection accuracy and has less computational complexity compared to other methods. Ablation experiments confirm the contribution of each component, and robustness is evaluated under H.264 compression and adversarial perturbation conditions.
The rapid advancement of the deep generative models has enabled the creation of highly realistic deepfake videos capable of manipulating facial content with high visual fidelity. Detecting such manipulations becomes more challenging when videos are stored or transmitted in compressed formats, as compression alters pixel-level artifacts and may obscure forensic traces. This paper proposes a compression-aware multimodal deepfake detection framework that jointly exploits spatial, frequency and motion information from compressed videos. The framework first extracts frames and localizes facial regions using the Multitask Cascaded Convolutional Network (MTCNN) detector. Spatial features are obtained from RGB frames using ConvNeXt, frequency-domain features are extracted from Discrete Cosine Transform (DCT) representations using a Swin Transformer and temporal motion patterns are captured from motion vectors using a 3D CNN combined with a Temporal Convolutional Network (TCN). In addition, codec-level artifacts such as motion vectors and DCT coefficients are utilized to capture compression-induced inconsistencies associated with manipulated content. The extracted features are integrated using an attention-based fusion module that adaptively weights modality-specific representations before classification. By jointly modeling visual and codec-domain information, the proposed framework achieves an accuracy (ACC) level of 99.63% and Area Under Curve (AUC) of 99.83% on the FaceForensics++ (FF++) dataset, improving robustness to compression artifacts and enhances deepfake detection performance across different compression levels.
Deepfake technology poses a critical threat to live video conferencing and biometric authentication. Existing detection models are either purely spatial—rendering them vulnerable to video compression—or rely on computationally heavy 3D-CNNs incompatible with strict real-time CPU latency constraints. We propose a highly optimized Two-Stream Spatio-Temporal architecture specifically designed for zero-latency live video evaluation. The framework extracts fine-grained spatial artifacts using a lightweight EfficientNet-B0 backbone, while a unidirectional Gated Recurrent Unit (GRU) models frame-to-frame physiological inconsistencies. Evaluated on a diverse 3,000-video corpus from FaceForensics++, Celeb-DF, and DFDC, the model achieved 98.21 cross-dataset accuracy and a 0.9978 AUC. Crucially, CPU inference requires only 125.63ms per 16-frame sequence—well below the 533ms threshold of a standard 30 fps camera—guaranteeing seamless, real-time overlay detection. Finally, the network's decision boundaries are mathematically validated using Explainable AI (XAI) activation maps and t- SNE clustering.
A detection framework is proposed that extracts per-video rPPG wave- forms via RhythmFormer and trains a suite of lightweight classifiers to distinguish real from synthesized physiologi- cal signals and shows that detec- tion difficulty is strongly method-dependent.
We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.
