Intelligent Phishing Detection System by Gabriel AdebayoIntelligent Phishing Detection System by Gabriel Adebayo

Intelligent Phishing Detection System

Portfolio project
Gabriel Adebayo

Intelligent Phishing Detection System 🔒

An end-to-end ML-powered solution for securing networks against phishing attacks.
A complete end-to-end Machine Learning project that detects phishing websites using supervised learning. It features a clean pipeline architecture, experiment tracking with MLflow & DagsHub, containerized packaging, and cloud-ready deployment.

🧭 Problem Statement

Phishing websites impersonate trusted brands to steal credentials and financial data. Rule-based filters struggle against fast-evolving attacks and novel domains.
Why AI solves this:
Learns patterns from large URL & page-level features (not just static rules).
Generalizes to previously unseen phishing sites.
Continuously improves with tracked experiments and retraining.

📦 Dataset

Name: Phishing Website Detection Dataset
Format: CSV
Features: 30+ web attributes (e.g., URL length, SSL state, domain age)
Target Variable: Result
-1: Phishing
1: Legitimate

🗂 Project Structure (Main Skeleton)

This is the main structure; the production codebase contains additional modules/configs beyond this outline.
networksecurity/
├── data/                 # Raw + processed data
├── notebooks/            # Jupyter notebooks for exploration
├── networksecurity/      # Core pipeline modules
│   ├── components/       # Data ingestion, transformation, training
│   ├── pipeline/         # Training & evaluation pipelines
│   ├── utils/            # Helper functions
│   ├── logger/           # Custom logging
│   ├── exception/        # Error handling
│   └── app.py            # (Optional) API layer for inference
├── config/               # Config files (YAML)
├── final_models/         # Trained models
├── mlruns/               # MLflow experiments
├── requirements.txt
├── main.py               # Pipeline runner
└── README.md

🧱 Features & Stack

Feature Technology Used Data Handling pandas, numpy, MongoDB Modeling scikit-learn (LogReg, RF), xgboost Tracking & Versioning MLflow (local/remote), DagsHub Pipeline Architecture Modular, OOP-based Deployment (API Service) app.py service, Dockerized, cloud-ready Model Evaluation Accuracy, Precision, Recall, F1-Score, ROC-AUC

🏗️ Infrastructure (Live Diagram)

Workflow Overview
Storage: AWS S3 (data, artifacts, models)
Tracking: MLflow (experiments, metrics) + DagsHub (remote tracking)
Deployment: Docker image → AWS ECR → AWS EC2 (serving)
Automation: GitHub Actions for CI/CD (build, test, tag, push)
Observability: Custom logger/ + exception/ modules

🚀 Live Pipeline Diagram

Loading
%%{init: {'theme':'default', 'themeVariables': { 'fontSize': '18px'}, 'logLevel': 'debug'}}%%
flowchart LR
    subgraph Dev[Developer Workflow]
      A[Code & Notebooks] --> B[Git Commit/Push]
    end

    B --> C[GitHub Actions CI/CD]
    C -->|Build & Test| D[Docker Image]
    D -->|Push| E[AWS ECR]
    E -->|Pull & Run| F[AWS EC2 Service]

    subgraph Data & Tracking
      G[(AWS S3: Data & Artifacts)]
      J[(MongoDB: Features & Logs)]
      H[MLflow Tracking Server]
      I[DagsHub Remote Tracking]
    end

    A --> G
    A --> J
    C --> G
    F --> G
    F --> J
    H <-->|Log Params/Metrics/Artifacts| C
    H --> I

    style G fill:#f6f8fa,stroke:#888
    style J fill:#e8f5e9,stroke:#47A248
    style H fill:#e3f2fd,stroke:#1e88e5
    style I fill:#fff3e0,stroke:#fb8c00

🔁 ML Pipeline Stages

Data Ingestion
Data Validation (schema, nulls, ranges)
Data Transformation (feature engineering, scaling/encoding)
Model Training (Logistic Regression, XGBoost, Random Forest)
Model Evaluation (metrics + artifacts logged to MLflow)
Model Pushing (saved locally and/or to AWS S3)
Loading
%%{init: {'theme':'default', 'themeVariables': { 'fontSize': '18px'}, 'logLevel': 'debug'}}%%
flowchart LR
    A[Data Ingestion: CSV, Database, API] --> B[Data Validation: Schema, Nulls, Ranges]
    B --> C[Data Transformation: Encoding, Scaling, Feature Engg.]
    C --> D[Model Training: ML Classifier]
    D --> E[Model Evaluation: Metrics, Validation]
    E --> F[Model Registry: MLflow Tracking]
    F --> G[Deployment: FastAPI + Docker]
    G --> H[Monitoring & Logging: Performance + Errors]



📒 MLflow Integration

Each pipeline step logs parameters, metrics, and artifacts to MLflow.
Runs can be tracked:
Locally → mlruns/ folder
Remotely → DagsHub (via MLflow env vars)
Facilitates model comparison, reproducibility, and rollbacks.

🧪 Model Performance (Illustrative)

Model Accuracy Precision Recall F1-Score ROC-AUC Logistic Regression 91% 90% 89% 89% 0.93 XGBoost 94% 93% 92% 92% 0.96 Random Forest 96% 95% 96% 95% 0.98
Why Recall matters: ⚠️ Missing a phishing site (false negative) is riskier than a false positive. 👉 Random Forest was chosen for the best Recall/F1 balance.

🧰 How to Run

# Clone repository
git clone https://github.com/yourusername/networksecurity.git
cd networksecurity

# Create & activate environment
python -m venv venv

# Windows
venv\Scripts\activate

# macOS/Linux
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Run the end-to-end pipeline
python main.py

# Inspect MLflow experiments locally
mlflow ui   # open http://127.0.0.1:5000

🔗 Remote Tracking (DagsHub)

Set MLflow environment variables for remote tracking
Skip mlflow ui when using DagsHub

☁️ Cloud & Deployment (Optional API)

Artifacts & Models: stored in AWS S3
Container Image: built with Docker, pushed to AWS ECR
Serving: pull & run on AWS EC2
CI/CD: automated with GitHub Actions
API: networksecurity/app.py exposes inference endpoints

🔮 Future Improvements

✅ Explainability (SHAP) → feature contribution & analyst trust
✅ Streamlit Dashboard → live insights & SOC analyst workflows
✅ Threat Intel Feeds → enrich predictions (OpenPhish / PhishTank)
✅ LLM-assisted Triage → natural language rationale for SOC teams
✅ Unit & Integration Tests → higher coverage & reliability
✅ Infra as Code → Terraform + AWS Secrets Manager

👤 Author

Adebayo Gabriel – ML Engineer (AI × Cybersecurity)

🔗 Links

⚡ This is more than a model—it's a production-minded AI system for real-world network security.

1 like · 10 views · Sep 1, 2025

Created an ML system to detect phishing websites using supervised learning.

Tags