Multimodal Vision Transformer Forensics for Deepfake Resistant Secure Authentication

The recent advent of extremely realistic deepfakes generated by GANs and diffusion models has presented an unsparing challenge to face-based authentication and remote identity verification. Tradition Single-mode detectors are lost during compression, after-processing, and to unseen generator families and thus hurt inauthentic that systems via imitation. In this paper, we propose Multimodal Vision-Transformer Forensics (MViT-Forensics), a transformer-based forensic framework luxurious for secure authentication applied in deepfake Scenario. MViT-Forensics processes spatial RGB embeddings (such as Swin/ViT backbone), frequency domain residuals (DCT/FFT process step) and physiological rPPG waveforms (attention-based temporal encoder) together, and combines them using a cross-attention multimodal transformer. The above anomaly-aware intrusion scoring head will generate a confidence score in the form of authenticity level for an authentication policy. We analyze the approach on four public benchmarks—FaceForensics++, Celeb-DF v2, DFDC Preview and the Multimodal DeepFake Dataset (MMDF)— in standard and compressed conditions. Our model reaches average accuracies of 98.1% (FF++), 96.4% (Celeb-DF), 94.7% (DFDC) and 97.2% (MMDF) with mean AUC$\approx 0: 99$, and maintains$>90$percent accuracy at high levels of compression while having better performance than strong CNN or transformer baselines. Ablations verify the complementary contributions in frequency and physiology branches and the superiority of cross-attention fusion. Our findings suggest that multimodal attention-based forensics with intrusion scoring could close the gap between academic detectors and operational authentication needs.