🔐 Resilient Forecaster

Defending LSTM Models Against Data Poisoning Attacks

---

📌 Overview

Resilient Forecaster demonstrates how machine learning models can be vulnerable to data poisoning attacks and how defensive mechanisms can restore model performance.

We simulate an attack on an LSTM-based stock price prediction system and measure:

• 📉 Performance degradation due to poisoning.
• 🛡 Recovery after applying statistical defense.

---

🎯 Problem Statement

Machine learning models rely heavily on clean training data. If an attacker manipulates training labels, the model learns incorrect patterns.

Can we detect and recover from data poisoning using automated defense?

---

🧠 Project Architecture

The workflow follows a modular pipeline:

1. Raw Stock Data: Ingested via yFinance.
2. Preprocessing: Scaling and sequence creation (Time-Series Windows).
3. Execution: Clean Training → Poisoned Training → Defended Training.
4. Comparison: Evaluation via Mean Squared Error (MSE).

---

⚙️ Technologies Used

• Python (Core Language)
• TensorFlow / Keras (Deep Learning)
• NumPy & Pandas (Data Manipulation)
• Scikit-learn (Preprocessing)
• Matplotlib (Visualization)
• yFinance (Stock Data)

---

🗂 Project Structure

Resilient_Forecaster/
├── src/
│   ├── attacker.py        # Poisoning logic
│   ├── defender.py        # Restoration logic
│   ├── model_lstm.py      # LSTM architecture
│   └── data_loader.py     # Data fetching
├── main.py                # Main pipeline
└── README.md              # Documentation