Deepfake Detection Pipeline
Eleven stages of forensic analysis — from raw frame to court-admissible verdict.
NuCoVault's backward-detection engine combines computer vision, signal forensics, and cryptographic provenance in a single pipeline. Every stage produces evidence. Every verdict is reproducible.
STAGE 01
Input Source
- Video / Image / URL
- Multi-format ingest
- Chain-of-custody stamp
STAGE 02
Preprocessing
- Sanitization
- Frame sampling
- Dedup — pHash / SSIM
STAGE 03
Detection Layer
- YOLO face detection
- Caffe fallback model
- Multi-face tracking
STAGE 04
Region Processing
- Bounding-box normalize
- Face crop
- ROI extraction
STAGE 05
Forensic Analysis Engine
- FFT high-freq ratio
- Laplacian variance
- Optical flow
- Chroma JS distance
- Residual noise
- PRNU-like correlation
STAGE 06
AI Model Fusion
- TinyDFNet CNN
- Heuristic fusion
- Optional ensemble
STAGE 07
Scoring Engine
- Frame-level probability
- Segment aggregation
- Temporal smoothing
STAGE 08
Decision Layer
- Thresholding
- Gate policy — OR / AND
- Confidence gating
STAGE 09
Output Verdict
- REAL / FAKE / UNCERTAIN
- Per-segment scoring
- JSON evidence report
STAGE 10
Attribution Engine
- Feature vector extraction
- MLP classifier
- Model / source attribution
STAGE 11
Provenance & Verification
- DCT-QIM watermark
- Gibbs signature
- Verification pipeline
- C2PA-compatible output
One pipeline. Eleven stages. A verifiable chain from pixel to provenance.