.png)
AI Generated Video Detection via Temporal Graphs and Neighborhood Consistency Regularization
“We propose a Graph Neural Network for detecting AI generated videos. By enforcing Neighborhood Consistency Regularization exclusively across temporal edges, our model prevents spatial overfitting to known artifacts. This structural constraint significantly improves out of distribution generalization, achieving state-of-the-art robustness on unseen generative models.”
Project Scope & Details
We propose a Graph Neural Network for detecting AI generated videos. By enforcing Neighborhood Consistency Regularization exclusively across temporal edges, our model prevents spatial overfitting to known artifacts. This structural constraint significantly improves out of distribution generalization, achieving state-of-the-art robustness on unseen generative models.
Spatiotemporal Graph Neural Networks
Models facial landmark spatial geometry and temporal optical flow dynamics across video frames as dynamic graphs.
Topological Inconsistency Detection
Identifies subtle deepfake rendering artifacts and temporal boundary warping invisible to traditional 2D CNNs.
Artifact Generalization Radar Evaluation
Systematically evaluates model robustness against cross-manipulation datasets, compression, and noise.
High-Throughput Video Verification
Accelerates graph convolution inference on Dual NVIDIA RTX 4090 GPUs to enable near-real-time authenticity screening.
Architecture Pipeline
Evaluation Matrix
- Mohit Jignesh Madhu
- Dr. Himangshu Sarma