TrustTrace

Banking-Grade Signature Detection and Verification Using Machine Learning and Computer Vision

TrustTrace is a state-of-the-art, multi-layered signature verification platform designed to secure financial ecosystems against advanced document forgery. Unlike traditional static template-matching systems, TrustTrace utilizes a dual-layered verification architecture combining robust static shape analysis, microscopic CV tremor inspection, and dynamic biometric tracking.

Overview

The financial and banking sectors rely heavily on signatures for high-value transactions. Manual verification is prone to fatigue and subjective bias, while standard machine learning models struggle to detect skilled tracings.

TrustTrace bridges these gaps by analyzing both the geometric structure of a signature and the physical mechanics of handwriting (such as micro-tremors, stroke continuity, ink pooling, and drawing velocity). Built with a high-performance FastAPI backend and an interactive Three.js-powered frontend dashboard, it provides real-time verdicts, Explainable AI heatmaps, and comprehensive transaction risk scores.

---

Key Features

1. Dual-Layered Verification

   • Static Analysis: Extracts Histogram of Oriented Gradients (HOG) features from scanned images.
   • Dynamic Tracking: LSTM/GRU temporal models analyze stroke coordinates, lifts, and speed variation for digital entries.

2. Soft-Voting Ensemble Model

   • Employs five classifiers: Support Vector Machine (SVM), Random Forest, Gradient Boosting, Logistic Regression, and K-Nearest Neighbors (KNN). This ensemble approach yields highly accurate spatial match verdicts.

3. OpenCV Micro-Tremor Engine

   • Calculates Edge Jitter Ratio to detect hand tremors typical of slow tracing.
   • Computes Ink Blot Ratio to identify hesitation pauses through localized pools of thick ink.

4. Prototypical Adaptive Learning (Few-Shot)

   • Monitors natural signature drift over time. It continuously updates the user's profile centroid in the latent space, avoiding the need for complete model retraining.

5. Explainable AI (XAI) Heatmaps

   • Generates pseudo Grad-CAM heatmaps highlighting abnormal stroke densities and anomalies, providing visual evidence to tellers.

6. Multimodal Financial Risk Engine

   • Fuses visual authenticity scores with contextual transaction metadata (amount, location, flags, account history) using a Bayesian weighting process to compute a final risk index.

7. Interactive Dashboard

   • Real-time HTML5 canvas pads, 3D particle animations, and detailed verification analytics in the teller UI.

---

The TrustTrace Advantage (Pros)

• Comprehensive Security: Defends against both casual forgeries (via shape analysis) and skilled traced forgeries (via tremor and rhythm analysis).
• Adaptable & Future-Proof: Automatically adjusts to changes in a user's handwriting over time without skyrocketing False Rejection Rates (FRR).
• Transparent Decisions: Does not act as a "black box." Visual heatmaps explain precisely why a signature is flagged.
• Context-Aware: Evaluates the risk of the transaction holistically rather than just looking at the signature in isolation.
• Accessible Deployment: Engineered to run efficiently without requiring heavy, expensive GPU infrastructure at every bank teller's desk.

---

How It Outperforms Existing Systems

Feature / Capability	Traditional Systems	Deep CNN Models	TrustTrace
Traced Forgery Detection	Poor (Foiled by overlays)	Moderate	Excellent (Micro-tremor CV)
Explanation of Results	None (Binary verdict)	Black-Box	High (Visual Heatmaps)
Signature Drift Handling	Static (High false rejects)	Requires Retraining	Dynamic (Adaptive Latent Centroids)
Context Integration	Isolated	Isolated	Integrated (Bayesian Risk Engine)
Hardware Requirements	Low	High (Dedicated GPUs)	Optimized (Runs locally via FastAPI)

---

Project Structure

• Frontend: HTML, CSS (style.css), JS (app.js) - Contains the Teller Dashboard and interactive elements.
• Backend/ML: Python pipeline in ml_pipeline/ - Includes the FastAPI service, Computer Vision core, and Machine Learning models.
• Documentation: VTU Project Reports (TRUSTTRACE_REPORT.*) detailing the methodology, architecture, and results.

Setup & Execution

Prerequisites

• Python 3.8+
• A modern web browser

Backend (ML Pipeline)

1. Navigate to the ml_pipeline directory:

   cd ml_pipeline

2. (Optional) Create and activate a virtual environment:

   # Windows
   python -m venv venv
   .\venv\Scripts\activate
   
   # Linux/Mac
   python3 -m venv venv
   source venv/bin/activate

3. Install the required dependencies:

   pip install -r requirements.txt

4. (Optional) Generate Data & Train the Model: Note: A pre-trained signature_svm_model.pkl is already included. You only need to run these if you wish to train the model from scratch.

   python generate_dummy_data.py
   python train.py

5. Start the FastAPI server:

   uvicorn main:app --reload

   The backend API will now be running at http://127.0.0.1:8000.

Frontend (Dashboard)

The frontend is built with plain HTML, CSS, and JS. It requires no complex build tools.

1. Open the Application: Simply double-click index.html in the root directory of the project to open it in your web browser.

   Alternatively, you can serve it via a simple local server:

   # In the root directory (where index.html is located)
   python -m http.server 3000

   Then navigate to http://localhost:3000 in your browser.

2. Usage: Ensure the FastAPI backend is running simultaneously so the dashboard can communicate with the ML pipeline to verify signatures.