🎣 PhishingDetector: Intelligent Anti-Bias URL Guard

PhishingDetector is a production-ready, end-to-end Machine Learning system designed to identify phishing URLs with high precision and transparency. Unlike traditional detectors that fall for "length-bias," this system features an Anti-Bias Engine optimized for complex e-commerce and search URLs.

Note

Versi Bahasa Indonesia tersedia di bawah. (Indonesian version below).

---

🚀 Key Innovations

1. Anti-Bias Engine (Feature Blindness)

Most phishing models incorrectly penalize long, complex URLs common in legitimate e-commerce sites (e.g., Tokopedia, Amazon). Our model employs Aggressive Feature Blindness, focusing exclusively on domain characteristics and verified security markers while ignoring path complexity. This improved legitimate URL detection from 0% to 90% in real-world complex scenarios.

2. Explainable AI (XAI)

Transparency is key to security. Every prediction is accompanied by a SHAP (SHapley Additive exPlanations) breakdown, showing exactly which features pushed the AI toward a "Phishing" or "Legitimate" decision.

---

🏗️ Technical Architecture

graph LR
    A[User Input URL] --> B[Feature Extractor]
    B --> C[Anti-Bias Preprocessing]
    C --> D[LightGBM Champion Model]
    D --> E[Inference Engine]
    E --> F[SHAP Explainer]
    F --> G[Real-time API & Dashboard]

Loading

Tech Stack

• ML Engine: LightGBM, Optuna (TPE Tuning), Scikit-learn.
• Explainability: SHAP (Local & Global Interpretation).
• Backend: FastAPI (Python), MLflow (Experiment Tracking), Uvicorn.
• Frontend: Next.js 15, React 19, Tailwind CSS v4, shadcn/ui.

---

📂 Project Structure

• backend/: FastAPI server and real-time inference logic.
• frontend/: Interactive dashboard with SHAP visualizations.
• src/mltools/: Custom library for standardized ML pipelines.
• models/: Optimized model artifacts and preprocessing pipelines.
• notebooks/: Comprehensive research and EDA process.

---

🛠️ Installation & Setup

Backend

# Install dependencies
pip install -r requirements.txt
pip install -e .

# Run API server
uvicorn backend.app:app --port 8001 --reload

Frontend

cd frontend
npm install
npm run dev

Visit http://localhost:3000.

---

🎣 PhishingDetector: Pelindung URL Anti-Bias Cerdas

PhishingDetector adalah sistem Machine Learning end-to-end siap produksi yang dirancang untuk mengidentifikasi URL phishing dengan presisi tinggi dan transparansi penuh. Dibangun untuk mengatasi kelemahan model tradisional yang sering terjebak dalam "bias panjang URL".

🚀 Inovasi Utama

1. Engine Anti-Bias

Banyak model phishing salah sangka terhadap URL panjang dan kompleks milik situs belanja online legal. Sistem ini menggunakan Aggressive Feature Blindness, yang memfokuskan deteksi pada karakteristik domain dan marker keamanan terverifikasi, mengabaikan kompleksitas path. Perbaikan ini meningkatkan akurasi pada URL kompleks dari 0% menjadi 90%.

2. Explainable AI (XAI)

Setiap prediksi disertai dengan visualisasi SHAP, menunjukkan secara transparan fitur mana yang membuat AI yakin bahwa sebuah URL adalah "Phishing" atau "Legit".

---

🛠️ Tech Stack

• ML Engine: LightGBM, Optuna, Scikit-learn.
• Interpretasi: SHAP.
• Backend: FastAPI, MLflow.
• Frontend: Next.js 15, React 19, Tailwind CSS.

---

📝 Creator

Fawwzrf AI Engineer & Full-stack Developer

---

Developed with precision for a safer web environment.