AI videos detected using first-digit gradient statistics and analysis
A Generalizable and Explainable Framework for Synthetic Video Detection Using First-Digit Gradient Statistics
Computer Vision and Pattern Recognition
Summary
Detecting AI-generated videos is getting harder because they can show many different scenes. The authors developed a new way to spot these videos by looking at the edges in the images using a special math rule about numbers, called the first-digit law. Their method doesn’t rely on knowing what kind of video generator made the video or details like compression, making it very general. They tested their approach on multiple big datasets and showed how it works even when others fail. This helps explain why their method can tell real from fake videos in many cases.
What this means in practice
- •For video security teams: Integrate the detection method to flag AI-generated videos regardless of scene or generator, enhancing video content authenticity checks.
- •For content moderation teams: Use the explainable gradient-based signals to improve automated detection of synthetic videos for safer online platforms.
Authors
Sidharth Shanu, Gautam Kumar, Tej Singh
Abstract
AI video generators have not only become harder to detect but are used to generate a diverse set of scenarios from landscapes to street views to animal videos. This creates a problem where CNN-based detectors are effective but offer no insight into their inner workings, while forensics-based detectors are often pretrained for a set scenario or become too complex to derive meaningful insights. We present a novel approach to AI video detection using Sobel gradient values analysed with the first-digit law. Using linear discriminant analysis, we visualise the discriminatory signal, while a multi-layer perceptron is used for classification. The detection method has no generator- or scenespecific features, and the model has no knowledge of container formats, codec, bitrate, or compression artefacts. The model is trained and tested on GenBuster-200K, GenBusterBench, GenVA, FaceForensics++ C23, and CelebDF. We also show how zero-shot detection fails even though the feature set carries a discriminatory signal.