An Effective CNN-Based Approach for Synthetic Face Image Detection in Pre-social and Post-social Media Context

The proliferation of image manipulation techniques, including DeepFake technology, has posed significant threats to the authenticity and credibility of images. Accurately classifying real and fake images has become crucial in fields such as forensics, security, and media authentication. However, detecting fake images during downloading and uploading from social networks is even more challenging. In this paper, we present an approach based on the EfficientNet model to learn discriminative features for classifying real and synthetic face images shared on social networks. We conducted extensive experiments using the TrueFace dataset, which comprises real and synthetic facial images shared on three major social media platforms. We employed the EfficientNet-B2 model trained on a combination of pre-social and post-social images from the TrueFace dataset. The presented approach outperforms all other methods, achieving accuracies of 99.98%, 100%, and 100% for images shared on Facebook, Telegram, and Twitter. This approach demonstrates exceptional performance when evaluated on a distinct dataset of images shared on social media platforms, separate from the images used for training.