Benchmarking Deep Learning Architectures for Deepfake Detection in Medical Images

With advanced digital manipulation techniques making it more challenging to distinguish between genuine and fake information, deepfakes pose a threat to the integrity of digital evidence within the health sector. This study reviews the performance of three deep learning architectures: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN), in the classification of medical chest x-ray images as real or fake. A balanced dataset of real and synthetic digital images is used to train the models. To support the generalisation ability of the models, operations such as image normalisation and augmentation are carried out on the images. Performance evaluation metrics such as accuracy, precision, recall, and F1-score are used to quantify and compare the performance of the models. The study found that CNN performs better than RNN and LSTM in detecting deepfake images. Spatial feature extraction is among the attributes that explain CNN's improved performance. RNN and LSTM models can handle sequential or temporal data, but they are challenged regarding the detection of spatial patterns in images. CNNs are also parameter efficient due to their high learning efficiency. CNNs prevent overfitting and improve generalisation by learning using fewer parameters, feasible through weight sharing and local receptive fields. The study outcome indicates that the CNN model was 97% accurate, outperforming the RNN model at 52% and the LSTM model at 54%. This demonstrates that CNNs are a reliable approach for deepfake image detection and recommends their adoption as the primary architecture for deepfake classification systems to preserve medical diagnostic integrity.