DeepScan AI uses EfficientNet-B4 as a deep feature extractor combined with 5 OpenCV-based computer vision signals to detect manipulated video content with high accuracy.
Every video is analyzed across multiple dimensions to ensure no manipulation goes undetected.
PyTorch EfficientNet-B4 (pretrained on ImageNet) extracts 1792-dimensional feature vectors per frame. Cosine similarity between consecutive frames detects GAN instability.
Laplacian operator measures face sharpness across frames. Deepfakes often have inconsistent sharpness — some frames blurry, some sharp — detected via coefficient of variation.
Sobel gradient operator detects blending seams at face boundaries. Face-swap deepfakes leave a sharp gradient discontinuity where the face was pasted onto the body.
LAB color space comparison between face region and background. Deepfake faces pasted from different sources have mismatched brightness and color channels.
Frame-to-frame pixel difference analysis on face crops. Real videos have smooth natural motion (CV 0.15–0.65). Deepfakes show erratic flickering (CV > 0.9).
All 5 signals are combined using weighted averaging (EfficientNet 45%, Texture 20%, Blend 15%, Color 15%, Flicker 5%). Score ≥ 38 → FAKE, < 38 → REAL.
A four-stage pipeline from upload to verdict in seconds.
Drag and drop or select any MP4, MOV, AVI, or WebM file up to 100MB.
Key frames are intelligently sampled and faces are detected and isolated for analysis.
EfficientNet-B4 extracts deep features and 4 OpenCV signals (Laplacian, Sobel, LAB color, Temporal) run on face crops.
Results are fused into a final REAL/FAKE verdict with full confidence breakdown and warning flags.
Upload any video file and get a full deepfake analysis report in seconds.
Supports MP4, MOV, AVI, WebM — up to 100MB
or click to browse files
Full breakdown of all signals and confidence scores