VIDS-guard: a novel forensics-aware multi-stream transformer framework for robust deepfake video detection

Abstract

The proliferation of highly realistic deepfake videos poses a growing threat to digital trust, underscoring the need for detectors that remain reliable across diverse manipulation types and capture conditions. This paper introduces VIDS-Guard (Video Integrity Deepfake Shield), a novel forensics-aware multi-stream transformer framework that integrates spatial, frequency, and temporal cues within a unified architecture. Unlike conventional convolutional or transformer-based detectors that rely primarily on semantic consistency, VIDS-Guard embeds forensic inductive biases through Spatial Rich Model (SRM) residual filtering, YCbCr color-space decomposition, and Fast Fourier Transform (FFT) spectral embeddings to expose subtle manipulation artifacts. A temporal transformer encoder with attention pooling further models cross-frame inconsistencies, enabling robust video-level predictions. Extensive experiments conducted with six benchmark models—Xception, ResNet50, MobileNetV3-Large, SlowFast, ViViT, and TimeSformer—demonstrate that VIDS-Guard achieves superior generalization and balanced detection performance across validation, test, and unseen datasets, attaining the highest accuracy and Macro-F1 scores under domain shift. These findings establish VIDS-Guard as a state-of-the-art forensic framework for trustworthy multimedia authentication and emphasize the importance of incorporating forensic priors to ensure sustainable robustness in deepfake video detection.