GenDet-M: A Multimodal Ensemble Architecture for Detecting High-Fidelity Synthetic Videos
Year: 2026
Type: article
Abstract: Over the past decade, diffusion models and generative adversarial networks have advanced to the point where synthetic videos are visually indistinguishable from real footage, enabling spread of misinformation, privacy violations, and fraud. Existing detectors rely primarily on unimodal spatio-temporal artifacts and generalize poorly to newer generators, particularly for multimodal video content where visual and audio cues must both be evaluated. We propose GenDet-M, a multimodal ensemble architecture that detects synthetic videos by extracting cross-modal artifacts across visual and acoustic domains. Three specialized streams handle this: a Transformer branch for temporal gradients, a 3D CNN branch for spectral and frequency-domain anomalies, and a BiGRU with self-attention for audio de-synchronization. Static concatenation fuses the modality streams independently and cannot model interactions between them; GenDet-M instead introduces a Context-Aware Multi-Modal Fusion module that first lets the streams exchange evidence through cross-modal attention and then applies a context-conditioned gate over the stream features. We evaluate GenDet-M on a custom dataset of high-fidelity multimodal videos produced by 2025 state-of-the-art generators (e.g., Veo3, Kling 2.6/2.1, and Seedance 1.0). On the held-out test split over five random seeds, GenDet-M attains 96.6 ± 1.2% accuracy, 99.1 ± 0.3% AUC, and 96.6 ± 1.2% F1 Score and significantly outperforms a static concatenation baseline (McNemar's test pooled over seeds, χ² = 9.13, p = 0.0025). Against six external detectors evaluated on our benchmark, GenDet-M is significantly more accurate (McNemar's test, p < 0.005 in every case), underscoring that detectors tuned on earlier generators do not transfer to 2025-generation synthesis.
Institutions: Leading University, Sylhet Engineering College
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225455151