GAN (Generative Adversarial Network)

The AI technique behind most deepfakes, where two AI models compete against each other until the fake becomes convincing enough to fool the other one.

One model acts like a forger and the other like an inspector, and each round makes the forger better at producing fakes that pass.

A machine learning architecture pairing a generative model, which produces synthetic media, against a discriminative model, which tries to detect it as fake, with the generator improving through repeated rounds until the discriminator can no longer reliably tell real from synthetic.

A generative adversarial network (GAN) trains in a loop:

1. The generator creates a synthetic image, video, or sound.
2. The discriminator examines it alongside real examples and guesses which is which.
3. The discriminator's feedback tells the generator how to improve.
4. The loop repeats until the fakes are very hard to distinguish from real media.

GANs were introduced in 2014 by Ian Goodfellow and colleagues. They are one of the main techniques behind deepfakes, and newer approaches, such as diffusion models, can also produce realistic images and video.

GANs sit next to terms that are easy to mix up:

• Deepfake is the synthetic media that a GAN can produce.
• Deepfake detection is the technology that tries to catch the output.
• Model drift is the decline in a model's accuracy as the patterns it was trained on change, which affects detectors facing new generation methods.

For more, see Deepfakes 101: Examples and deepfake fraud prevention tips.