DIDM: Deepfake Image Detection Model Using Deep Learning

This paper presents a structured approach using fundamental tools and techniques for detecting fake images using deep learning (DL). The dataset contains images which are trained, tested and validated. The test sets undergo preprocessing, including data augmentation and rescaling by freezing and unfreezing layers for consistency. EfficientNetB4, a pre-trained Convolutional Neural Network (CNN) model on ImageNet, serves as the backbone, with its base layers frozen and additional layers added for binary classification. Training employs learning rate scheduling for optimization techniques like ModelCheckpoint, EarlyStopping, TensorBoard, LearningRateScheduler and ReduceLROnPlateau. ImageDataGenerator and CutMiz are used for data augmentation, whereas focal loss is evaluated for generating the loss function. This paper uses some data from FaceForensics++. Further, unfreezing of 300 layers with a reduced learning rate to enhance performance is done by fine-tuning, and AdamW optimizer is used for the optimisation of the code, which helps to ensure less overfitting and is found to be useful during finetuning of large, trained model. The evaluation metrics, like accuracy, loss, precision, recall and F1-score are also predicted. The final model is a combination of EfficientNetB4 and Custom Neural Network head (Attention Mechanism, Global Average Pooling, Global Max Pooling, Dense Layers with Swish Activation & Dropout, Final Dense Layer with Sigmoid Activation) is further evaluated on test data and is applied to make probability-based classifications, ensuring robustness and real-world applicability in image authentication.