DOCUMENT FORENSICS & DEEPFAKE DETECTION AT SCALE
Built for a global identity-verification provider
An identity-verification platform classifies ID documents, extracts their data and runs forensic and deepfake checks before returning a risk decision. The pipeline combines computer vision, document intelligence and custom tamper analysis inside a mobile-first review experience.
By Yashwant Daniel · Solvstrat
Key
Facts
- Sector
- Identity verification
- Built
- ID forensics and deepfake detection pipeline
- Stack
- YOLO · Azure Document Intelligence · PyTorch · React
- Constraint
- Real-time decisions with explainable evidence
- Outcome
- Automated document triage with forensic audit trail
Challenges
Fraud teams were reviewing identity documents manually while synthetic and tampered documents became cheaper to produce. The platform needed to classify documents, extract data and detect forgery at volume without turning every check into a human review.
Document variety
Government IDs differ by country and issuer in layout, language, security features and data fields.
Tampering and synthetic media
Edited images, forged documents and deepfake video could pass a simple visual check.
Manual review load
Every uncertain document became a human task, limiting throughput and slowing onboarding.
Constraints
Decisions must return in real time rather than through overnight batch review.
Every flagged document must carry explainable forensic evidence.
The review experience must work on mobile devices first.
All decisions and evidence must be retained for audit.
The pipeline must handle multiple document types without bespoke flows per country.
Reference
Architecture
Capture and classification
The client captures the document, and a YOLO classifier identifies its type. Classification first prevents the wrong extraction model from being applied.
YOLO · Document capture
Extraction
Azure AI Document Intelligence and custom field models read the document and return structured values for verification against expected data.
Azure AI Document Intelligence · Custom models
Forensic analysis
Checks examine metadata, compression patterns, pixel-level anomalies and digital tampering signals to determine whether the document was altered.
OpenCV · Image forensics
Deepfake and liveness
Model-based checks look for synthetic faces, replay attacks and injected media that would otherwise defeat a document-only review.
PyTorch · Liveness models
Decision and review
Results are combined into a risk decision. Clear cases pass automatically; flagged cases enter a mobile-first review queue with evidence attached.
Risk rules · Review UI · Audit store
Data
Flow
- 1
The user captures an identity document on a mobile device.
- 2
The classifier identifies the document type.
- 3
The extraction model returns structured fields.
- 4
Forensic checks analyze integrity and tampering signals.
- 5
Deepfake and liveness checks evaluate the person and media.
- 6
A risk decision is returned, and flagged cases are queued with evidence.
Delivery
Classification and extraction
Document typing and structured field extraction.
Forensics
Tampering, integrity and image-forensic checks.
Deepfake and liveness
Synthetic-media and replay detection.
Review experience
Mobile-first review queue and decision evidence.
Hard
Problems
Document variety
Problem: A single extraction model degraded badly across issuers and languages.
Resolution: Classification runs first, and extraction is selected per document type, so each model handles a narrower, more reliable problem.
False positives
Problem: Aggressive tampering checks flagged legitimate low-quality captures.
Resolution: Signals are combined rather than treated individually, and uncertain cases are routed with evidence instead of auto-rejected.
Latency at volume
Problem: Running every check sequentially made real-time decisions impossible.
Resolution: Independent checks run concurrently, and only the required checks are invoked for each document type.
Results
Automated triage
Clear documents pass without manual review; uncertain cases are routed with their evidence attached.
Explainable decisions
Each result carries the checks and signals that produced it, so reviewers can see why a document was flagged.
Mobile-first review
Capture and review work on small screens with a responsive breakpoint strategy.
Technology
Used
YOLO
Azure AI Document Intelligence
PyTorch
OpenCV
React
PostgreSQL
Lessons
Classify before you extract — it is the cheapest accuracy improvement in the pipeline.
A flagged result without evidence is just a rejection; evidence is what makes review possible.
False-positive handling is a product decision, not a model metric.
CLIENT SUCCESS STORIES
More Case Studies
Explore how Solvstrat engineers custom digital architectures and software platforms for leading brands.
FAQ
Frequently asked questions
Document forensics is the process of verifying that an identity document is genuine and unaltered. It examines metadata, compression and pixel-level artifacts, and checks extracted data against expected values. It is different from OCR, which only reads text.
Deepfake and liveness checks confirm that a real, present person is behind the document and selfie. They look for synthetic faces, replays and injected media before an account is created.
Yes. Each decision carries the checks and signals that produced it, so a human reviewer can inspect the evidence rather than trusting an opaque score.
The review experience is mobile-first, with a responsive breakpoint strategy so capture and review work on small screens as well as desktop.
SEE THE PROTOTYPE WE'D BUILD
Give us 30 minutes.
We'll show you the fix.
Tell us where the work slows you down. We'll come back with a prototype of the custom software, internal tools or automation we'd build to remove it — working in weeks, and you own the code.