DeceptiScan 🛡️🔍

DeceptiScan is a premium, AI-powered browser extension and machine learning platform designed to detect fake, AI-generated, and manipulative product reviews on e-commerce sites in real-time. By bridging a React + Plasmo extension frontend with a FastAPI + PyTorch (DeBERTa-v3) inference backend, DeceptiScan calculates an actionable Trust Score, details review patterns, and provides explainable AI insights for safer shopping.

---

🌟 Key Features

💻 Premium Dashboard

• Intuitive UI: A sleek dark/light theme-adaptive pop-up window built with glassmorphism design principles.
• Trust Score Gauge: Visualizes seller/product reliability using an animated circular gauge.
• Tabbed Analytics:
  • Overview: Summary of scraped reviews, product metadata, and final sentiment rating.
  • Reviews: Labeled feed of reviews (marked as Genuine, Suspicious, or Highly Fake) with prediction confidence percentages.
  • Signals: Breakdowns of warning triggers (such as review count discrepancies or abnormal distributions).
  • Settings: Adjustable connection settings (e.g., configuring local/remote server endpoints).

🕷️ Smart Scraper (Anti-Bot & Dynamic)

• Deep Scrape: Asynchronously crawls up to 5 paginated pages of reviews directly from Amazon product listings.
• Anti-Detection Jitter: Employs randomized delay intervals (700ms - 1500ms) and cookie-integrated fetch requests to query pages without triggering CAPTCHAs or IP blocks.
• DOM Parsing: Parses raw HTML dynamically without relying on brittle selector dependencies, collecting reviewer details, ratings, titles, and body texts.

🧠 Fine-Tuned NLP Classification Backend

• Primary Classifier: Houses a fine-tuned DeBERTa-v3 sequence classification transformer model optimized on over 72,000 labeled review datasets.
• Multi-Core & GPU Acceleration: Automatically routes inference tasks to PyTorch CUDA cores for lightning-fast inference (< 200ms per batch).
• Explainable Trust Score: Weighs classification confidence alongside rating distributions and suspicious reviewer behaviors to output a single, comprehensive percentage.

---

🏗️ System Architecture

graph TD
    subgraph Browser Extension [Frontend & Scraper]
        A[Amazon Product Page] -->|Injects| B(amazon.ts Content Script)
        B -->|Background Scrape| C{DOM Parser & Fetcher}
        C -->|Reviews JSON| D[popup.tsx Dashboard]
        D -->|CORS HTTP POST /analyze| E[FastAPI Gateway]
    end

    subgraph Backend Server [Inference API]
        E -->|Processes Request| F[Analyzer Service]
        F -->|Batch Inputs| G[DeBERTa-v3 Transformer Model]
        G -->|Predictions & Confidences| H[Trust Score Engine]
        H -->|Generates Report| F
        F -->|JSON Response| D
    end

Loading

---

📂 Project Structure

Deceptiscan/
├── .github/                       # GitHub actions pipelines
│   └── workflow/                  # CI/CD test automation for backend and extension
├── LICENCE                        # MIT License
├── README.MD                      # Project Documentation
├── github_upload_guide.md         # Upload & Repository Management Guide
└── extension/                     # Chrome Extension Workspace
    ├── assets/                    # Static image resources and extension icon
    ├── contents/                  # Content scripts (Amazon scraper injection)
    ├── lib/                       # Utility scripts & API request handlers
    ├── popup.tsx                  # Main dashboard React popup UI code
    ├── tsconfig.json              # TypeScript compilation setup
    └── backend/                   # FastAPI NLP Inference API
        ├── app/                   
        │   ├── main.py            # FastAPI Entry Point (Uvicorn Gateway)
        │   ├── api/routes/        # Endpoint routes (Analyze, Health, Reviews)
        │   ├── core/              # Global configurations & constants
        │   ├── schemas/           # Pydantic Request & Response models
        │   └── services/          # Inference orchestration & Trust Score calculations
        ├── training/              # Training scripts and evaluation Jupyter Notebooks
        ├── requirements.txt       # Python Dependencies list
        └── s.DockerFile           # Backend container build instructions

---

🚀 Getting Started

Follow these instructions to install, build, and run the complete DeceptiScan suite locally.

Prerequisites

• Node.js (v18 or higher recommended)
• Python 3.10+
• Google Chrome (or any Chromium-based browser)

---

Step 1: Install & Launch the Extension (Frontend)

1. Navigate into the extension directory:

   cd extension

2. Install dependencies:

   npm install

3. Start the Plasmo developer development server:

   npm run dev

4. Open Google Chrome and go to chrome://extensions/.
5. Enable Developer mode using the toggle switch in the top-right corner.
6. Click Load unpacked in the top-left and select the directory: Deceptiscan/extension/build/chrome-mv3-dev
7. The DeceptiScan icon will now appear in your extension toolbar!

---

Step 2: Set Up & Launch the Inference Server (Backend)

1. Navigate to the backend directory:

   cd extension/backend

2. Create and activate a virtual environment:

   python -m venv .venv
   
   # On Windows (PowerShell/CMD)
   .venv\Scripts\activate
   
   # On macOS/Linux
   source .venv/bin/activate

3. Install the backend Python dependencies:

   pip install -r requirements.txt

4. Download and Place Model Weights: Since model weights are not uploaded to GitHub due to size limitations, ensure your fine-tuned model folder deberta-v3-detector is placed in: extension/backend/models/deberta-v3-detector/ (Ensure it contains model.safetensors, config.json, tokenizer.json, etc.)
5. Start the FastAPI Server:

   python app/main.py

   The server will startup at http://localhost:8000. You can test it by visiting the Swagger Docs at http://localhost:8000/docs.

---

📊 Dataset & Model Details

The model was fine-tuned on a merged, deduplicated, and cleaned corpus of e-commerce reviews:

• Total Labeled Samples: 72,877
• Training Split: 57,325 reviews
• Testing Split: 14,332 reviews
• Primary Architecture: DeBERTa-v3 base sequence classifier

Performance Metrics

• Baseline XGBoost Accuracy: ~84.2%
• Fine-Tuned DeBERTa-v3 Accuracy: ~96.8%
• Average Batch Inference Latency: ~12ms per review (on CUDA GPU)

---

🛡️ License

Distributed under the MIT License. See LICENCE for details.