Disinformation at Scale: Detecting AI-Human Composite Images via Convolution Ensembles

Detecting fake news and synthetic content has become increasingly crucial in the cybersecurity community due to the rapid spread of disinformation, which can manipulate public opinion, influence elections, and harm reputations. The rise of advanced image generation models, like diffusion models, has exacerbated this issue by creating highly realistic images that are difficult to distinguish from authentic ones. Additionally, AI-human composite images, images with both human-generated and AI-generated elements, are more difficult to detect for modern detection tools. This paper explores state-of-the-art tools for detecting AI-generated images and introduces a novel detection method using ensembles of convolutional neural networks. Our method outperforms existing techniques when detecting AI-human composite images by dividing images into multiple subsections and evaluating each independently for signs of manipulation or generative signatures. We demonstrate the effectiveness of our approach through rigorous experimentation, showing significant improvements in detecting AI-human composite fake images and enhancing the robustness of detection systems against common evasion techniques like JPEG compression and geometric downsampling.