The Technology Behind Deepfakes

Deepfakes are produced by machine learning models that learn to synthesize a person's appearance from reference data. Two main architectures dominate: Generative Adversarial Networks (GANs) and diffusion models. GANs use two competing neural networks, a generator that creates images and a discriminator that tries to identify fakes, locked in a training loop until the generator produces results the discriminator cannot distinguish from real photographs.

Why Deepfakes Have Become So Accessible

In 2026, the computational cost of deepfake generation has dropped by over 95% compared to 2020. Consumer-grade graphics cards can run deepfake pipelines locally. Open-source models are freely available. Dedicated "nudify" and face-swap applications, many targeting non-consenting individuals, operate as subscription services accessible from any browser.

What Data Deepfakes Require

A convincing face swap requires as few as 15 to 30 clear reference photographs. Social media profiles, professional headshots, and publicly indexed images are sufficient training data. This means anyone with a public online presence is technically vulnerable. The more reference images available, the more realistic the output.

Legal Consequences in 2026

Creating or distributing non-consensual intimate deepfakes is a federal offense under the TAKE IT DOWN Act. State-level laws in 49 jurisdictions add additional criminal and civil liability. ScanErase uses biometric scanning to detect deepfakes regardless of how heavily they have been re-encoded, cropped, or filtered.