Synthetic Image Detection Via Supervised Contrastive Learning

Generative models like Diffusion Models (DMs), Variational Auto-Encoders (VAEs), and Generative Adversarial Networks (GANs) produce realistic synthetic images, posing a challenge for identifying manipulated or misleading content. To address this, we introduce ImagiNet, a high-quality dataset designed to mitigate biases and improve synthetic image detection across diverse content types. Utilising ImagiNet, we train a ResNet model with a self-supervised contrastive (SelfCon) learning approach, demonstrating state-of-the-art performance and speed on existing benchmarks, even under challenging social network conditions (resize and compression), making it a practical tool for dealing with the spread of misinformation and manipulated media.