AI Generated Media Detection: A Multimodal Ensemble Architecture to Combat Synthetic Media

Abstract

Over the past decade, diffusion models and generative adversarial networks haveadvanced to the point where synthetic videos are visually indistinguishable fromreal footage, enabling spread of misinformation, privacy violations, and fraud.Existing detectors rely primarily on unimodal spatio-temporal artifacts and gen-eralize poorly to newer generators, particularly for multimodal video contentwhere visual and audio cues must both be evaluated. We propose GenDet-M,a multimodal ensemble architecture that detects synthetic videos by extract-ing cross-modal artifacts across visual and acoustic domains. Three specializedstreams handle this: a Transformer branch for temporal gradients, a 3D CNNbranch for spectral and frequency-domain anomalies, and a BiGRU with self-attention for audio de-synchronization. Static concatenation fuses the modalitystreams independently and cannot model interactions between them; GenDet-Minstead introduces a Context-Aware Multi-Modal Fusion module that first letsthe streams exchange evidence through cross-modal attention and then appliesa context-conditioned gate over the stream features. We evaluate GenDet-M ona 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-outtest 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 outperformsa 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 everycase), underscoring that detectors tuned on earlier generators do not transfer to2025-generation synthesis