Autoencoder-Based Anomaly Detection (Credit Card Fraud Dataset)

🚀 Project Overview

This project implements an Autoencoder-based Anomaly Detection system using the Credit Card Fraud Detection dataset. The model is trained only on normal transactions and learns to reconstruct them with low error. Fraud samples (anomalies) produce a high reconstruction error, allowing the system to detect anomalies effectively.

---

🎯 Objectives

• Build a fully functional encoder–decoder (Autoencoder) from scratch.

• Learn representation learning and anomaly detection fundamentals.

• Use reconstruction error as the anomaly score.

• Evaluate with:

  • ROC-AUC
  • Precision, Recall, F1
  • Confusion matrix

• Provide an inference pipeline to score any transaction.

---

📂 Project Structure

project-autoencoder/
  ├── data/
  │     └── creditcard.csv
  ├── models/
  │     ├── autoencoder_best.pth
  │     ├── scaler.pkl
  │     ├── X_test_scaled.npy
  │     └── y_test.npy
  ├── notebooks/
  │     ├── 01_eda_and_preprocessing.ipynb
  │     ├── 02_training_logs_and_visuals.ipynb
  │     └── 03_anomaly_demo.ipynb
  ├── src/
  │     ├── model.py
  │     ├── train.py
  │     ├── eval.py
  │     └── inference.py
  ├── requirements.txt
  └── README.md

---

🧠 What is an Autoencoder?

An Autoencoder is a neural network with two parts:

1. Encoder → Compress input to a small latent vector
2. Decoder → Reconstruct input from latent

When trained only on normal data:

• Normal samples → low reconstruction error
• Fraud samples → high reconstruction error

This makes it perfect for anomaly detection.

---

📦 Installation

1. Clone the repository

git clone https://github.com/Tanish-Sarkar/Autoencoder_from_Scratch
cd Autoencoder_from_Scratch

2. Create and activate a virtual environment

Windows:

python -m venv venv
venv\Scripts\activate

Mac/Linux:

python3 -m venv venv
source venv/bin/activate

3. Install requirements

pip install -r requirements.txt

---

📊 Training the Autoencoder

Option 1: Train via notebook

Open:

notebooks/02_training_logs_and_visuals.ipynb

and run all cells.

Option 2: Train via script

Run:

python -m src.train

This will save:

• Best model → models/autoencoder_best.pth
• Scaler → models/scaler.pkl
• Test data → models/X_test_scaled.npy, models/y_test.npy

---

🧪 Evaluation

Run:

python -m src.eval

This prints:

• ROC AUC
• Best threshold (based on F1)
• Confusion matrix
• Precision, recall, F1
• Number of predicted anomalies

📈 Example Output

ROC AUC: 0.9470
Best F1 threshold: 0.588957
Confusion Matrix:
[[56768    95]
 [  129   363]]
F1 Score (fraud class): 0.7642

---

⚡ Inference (Score New Samples)

Run:

python -m src.inference

This demo:

• Loads trained model
• Scores a few test samples
• Outputs reconstruction error and predicted anomaly flag

Use inside other Python code:

from src.inference import score_samples
import numpy as np

sample = np.array([...])  # shape: (30,)
errors, flags = score_samples(sample)
print("Error:", errors[0])
print("Anomaly:", flags[0])

---

📑 Key Results

• ROC AUC: ~0.94
• Fraud F1 Score: ~0.76
• High precision + strong recall for anomaly class
• Very low false positive rate
• Strong evidence Autoencoder learned normal patterns well

These results align with literature on credit card anomaly detection.

---

🧱 Model Architecture

Encoder

• Linear → 128 → ReLU
• Linear → 64 → ReLU
• Linear → latent_dim

Decoder

• Linear → 64 → ReLU
• Linear → 128 → ReLU
• Linear → input_dim

---

📘 Tech Stack

• Python
• PyTorch
• NumPy / Pandas
• scikit-learn
• Matplotlib / Seaborn
• Jupyter Notebooks

---

🏁 Next Steps / Future Work

• Add Variational Autoencoder (VAE)
• Add latent space visualization (t-SNE, UMAP)
• Wrap inference in FastAPI
• Deploy endpoint for real-time anomaly scoring
• Train deeper AE for better fraud recall

---

👤 Author

Tanish Sarkar (Data Science & Machine Learning Enthusiast)