Initializing...
Loading forensic verification components...
// SYSTEM ARCHITECTURE RATIONALE
Why a Tripartite Architecture?
Real-world identity fraud is highly sophisticated, as demonstrated in the Election Commission's West Bengal Special Intensive Revision (SIR) where over 30 lakh voter records were flagged for anomalies. Fraud vectors are multi-modal: they range from advanced generative AI (like Google Gemini-generated voter ID cards) to logical database contradictions (such as birth certificates issued before the birth date, or dead electors listed as active). A single model or modality is easily bypassed. Decomposing the validation space into three orthogonal domains—Spatial, Frequency, and Semantic—eliminates single points of failure, ensuring that evasion in one stream is caught by the others.
Addressing Multi-Modal Fraud Signals
- Spatial Forensics: Specifically designed to combat AI-synthesized identity cards (like the Gemini-generated EPIC cards found in Bhangar). EfficientNet-B4 captures local text/seal aberrations, while the Vision Transformer (ViT) identifies long-range structural deviations from generative models.
- Frequency Forensics: Exposes local digital editing (such as altering dates on West Bengal birth certificates). Local edits require re-saving segments, leaving telltale JPEG block quantization mismatches and upsampling grid periodicities in FFT/DCT space.
- Semantic Forensics: Directly flags logical anomalies that humans miss (e.g. birth certificates issued prior to the birth date, impossible 2-to-4-month age gaps between siblings, UPI QR code mismatch, or ICDS maternal-welfare certificates issued to male electors).
calibrated late-stage fusion
KYC and electoral security require transparent audit trails. Monolithic end-to-end models act as opaque black boxes and are prone to false positives (such as approving dead voters or blocking legitimate citizens due to minor scanner noise). By running independent parallel streams and fusing their soft probability scores using a calibrated ensemble of XGBoost + LightGBM fortified by Isotonic Regression, we achieve auditable risk classification (e.g., flagging the 98.4% fraud probability seen here) that micro-observers and KYC teams can visually inspect.