Detecting-Insurance-Fraud

Introduction

Given images of damaged cars, both AI generated and real accidents, build a classifier that can identify an actual accident from a fraudulent one.

Setup and Dependencies

• TensorFlow 2.x
• Keras
• scikit-learn
• matplotlib
• seaborn

Data

The data can be found here: https://drive.google.com/drive/folders/17I3-H0PTz2Vh3liqIUOXOVjBmD9KJTXG?usp=sharing

Data Preprocessing

• ImageDataGenerator is used for data augmentation and preprocessing.
• Images are resized to 224x224 pixels.
• Training and validation data are split with a validation split of 15%.

Model Architecture

• The VGG16 model is loaded with pretrained ImageNet weights.
• The top layers of the VGG16 model are removed, and additional Dense layers are added.
• GlobalAveragePooling2D is used to flatten the output of the base model.
• LeakyReLU activation, Dropout, and BatchNormalization are applied to Dense layers.
• The output layer consists of a single neuron with a sigmoid activation function.

Training

• Adam optimizer with a learning rate of 0.001 is used.
• Binary crossentropy loss is used as the loss function.
• Metrics tracked during training include accuracy, precision, and recall.
• ModelCheckpoint and EarlyStopping callbacks are used for saving the best weights and early stopping.

Models

The fitted models can be found here: https://drive.google.com/drive/folders/1LpoJszTF6ya_4MmiIm3EJCDN1Tnl_PjL?usp=sharing

Evaluation

• Model performance is evaluated on a separate test set.
• Accuracy, Precision, and Recall metrics are calculated.
• Classification report and confusion matrix are generated to evaluate model performance.

Results

• The model achieves an accuracy of 93.4%, precision of 93.4%, and recall of 100% on the test set.

Conclusion

The VGG16-based image classification model demonstrates effective performance in distinguishing between Fraud and Non-Fraud classes.