Authenticating AI-Generated Social Media Images Using Frequency Domain Analysis.

Living in the age of social media, it is a daily routine for individuals to post videos, audio, pictures, and text online. In addition, the proliferation of Artificial Intelligence (AI) technology allows customizing multimedia content to meet personal demands. However, the popularity of AI-based text-to-image generators like DeepAI also opens the door to generating images for social media platforms that impersonate unsuspecting users without their permission. While people enjoy high creativity, such “fake” images could enable the propagation of deceptive information that negatively impacts an individual's personal life and potentially cause public unrest. Therefore, reliable methods to facilitate image authentication are vital to identify and flag them. In this paper, we present AUSOME-2, an upgraded version of our system that AUthenticates SOcial MEdia images (AUSOME) using frequency analysis technologies and machine learning (ML) algorithms. Images from several text-to-image platforms, such as Dall-E 2 and Google Deep Dream, are distinguished from genuine images. Spectral analysis techniques are used to obtain features and fingerprints in the frequency domain. These features enable the ML model to classify AI -generated social media images from genuine ones. The experimental results, on top of a proof-of-concept prototype, showed that the AUSOME-2 system is a promising approach to authenticate images with decent detection accuracy.