Detecting Deception: A Comparative Study of CNN Approaches for Deepfake Image Detection

The exponential growth of Artificial Intelligence (AI) has given rise to deepfake images, which present a major impact on digital safety, mistaken identity, and the spread of fake information. Due to their highly realistic nature, detecting deepfakes has become a major challenge, highlighting the critical need for robust detection systems. This research investigates the performance of nine advanced Convolutional Neural Network (CNN) architectures, namely: Custom CNN, ResNet50, VGG16, InceptionV3, DenseNet, MobileNetV2, XceptionNet, MesoNet, and InceptionResNetV2 which were used to detect deepfakes. To evaluate the effectiveness of the above CNN architectures, two large-scale image datasets were used: the Real & Deepfake Dataset and the 140K Real and Fake Faces Dataset. These datasets include more than 2,00,000 real and fake images. The training was done using a GPU to increase the speed and accuracy of the models. Among all the models, the Custom CNN provided the best results, reaching 99% accuracy on the Real & Deepfake Dataset and 98% on the 140K dataset. ResNet50, DenseNet, and InceptionV3 also performed very well, each scoring above 90% accuracy in both datasets. Other models like VGG16, MobileNetV2, XceptionNet, and InceptionResNetV2 also showed strong results, proving they are useful for deepfake detection. This analysis study aims to identify the most efficient and accurate CNN architecture for image-based deepfake detection, ensuring digital content safety, and identifying the fastest model for real-time deepfake detection in future applications.