Fraud Detection System

This project implements a fraud detection system using an autoencoder for feature extraction and a dense neural network for classification. The dataset used in this project contains transactions from credit cards; it is specifically designed to detect fraudulent transactions.

Prerequisites

• Python 3.6+
• pandas
• numpy
• scikit-learn
• keras
• TensorFlow
• imbalanced-learn
• joblib

You can install the necessary libraries using pip:

pip install pandas numpy scikit-learn keras tensorflow imbalanced-learn joblib

Dataset

The dataset named creditcard.csv should be structured with transaction features labeled from V1 to V28, a Time feature indicating the seconds elapsed between each transaction and the first transaction in the dataset, an Amount feature which is the transaction Amount, and a Class feature where 1 stands for fraudulent transaction and 0 stands for non-fraudulent.

Feature Engineering

Advanced feature engineering steps include:

• Log transformation of the Amount to reduce skewness and scaling issues.
• Extracting the hour of the day from the Time feature, which might capture intra-day patterns.
• Interaction feature between the log-transformed amount and the hour to capture any specific patterns at different times of the day.

These features replace the original Time and Amount features in the dataset.

Model Architecture

Autoencoder

• Input Layer: Size equal to the number of input features (after preprocessing).
• Encoder: Dense layers with decreasing units, regularizers, and dropout to learn compressed representation.
• Decoder: Mirrors the encoder architecture to reconstruct the input features.

Classifier

• Input Layer: Compressed feature set from the autoencoder.
• Hidden Layers: Dense layers with dropout to prevent overfitting.
• Output Layer: Single neuron with sigmoid activation to output the probability of fraud.

Training

• The autoencoder and classifier are trained using Adam optimizer with early stopping based on validation loss/auc.
• The datasets are first split into training and test sets. The training data is oversampled using SMOTE to handle class imbalance.
• Features are scaled using RobustScaler to handle outliers effectively.

Evaluation

• The model's performance is evaluated using ROC-AUC, precision-recall AUC, accuracy, confusion matrix, and classification report metrics.
• Predictions are made on the test set, and evaluation metrics are computed to assess the performance.

Usage

The trained model and scaler are saved to disk. For new predictions:

1. Load the saved models and scaler.
2. Apply the same preprocessing and feature engineering steps to the new data.
3. Predict using the classifier.

Saving and Loading Models

Models and the scaler object are saved using model.save() and joblib.dump(), respectively. This allows for easy distribution and deployment of the trained system for later use without retraining.