Comparative Review and Initial Implementation of GAN and Diffusion Model Artifacts for Deepfake Detection

Today, it's getting nearly impossible to tell what's real online. Hyper-realistic fake media (deepfakes), churned out by advanced AI like GANs and the newer Diffusion Models (DMs), are seriously undermining our trust in photos and videos .The big problem for security experts is the "generalization gap." Our detection tools learn to spot fakes from, say, a GAN, but then they completely fail when faced with a fake from a DM. It's an AI arms race where our defenses are always a step behind .This paper tackles that challenge head-on. We compare the unique digital "fingerprints" left behind by these two major AI families. Our technical approach uses Frequency Domain Analysis—a fancy way of looking at the noise patterns in images, specifically using the Power Spectral Density (PSD)—to measure and compare real images against both GAN and DM fakes .Our findings clearly show that the AI models are not faking things the same way: GANs leave distinct periodic noise patterns, while DMs leave unique high-frequency irregularities. These differences are measurable and significant. By quantifying these non-uniform flaws, we provide the essential evidence needed to finally build detection tools that can spot a fake, no matter which modern AI created it.