Generative Adversarial Network
A generative adversarial network is a pair of neural networks trained in competition: a generator that produces synthetic data from random input and a discriminator that tries to tell synthetic from real, each improving against the other until the generator's output is hard to distinguish from real data. Introduced in 2014, GANs produced the first photorealistic synthetic faces and the first deepfakes, led image generation until diffusion models overtook them around 2022, and remain in use for synthetic data.
- Also known as
- GAN, GANs, Adversarial training (generative)
- Used at
- image synthesis (StyleGAN's faces at thispersondoesnotexist.com), face-swapping deepfakes, image enhancement, synthetic data for training and testing (including synthetic transactions for fraud-model development), and as a conceptual model for adversarial fraud — the fraudster as generator, the fraud model as discriminator
- Binding
- no direct regulation; deepfake and synthetic-content rules (the EU AI Act's transparency obligations) apply to the outputs regardless of method
- Typical mechanics
- the generator maps random noise to an output; the discriminator classifies real versus generated; the two are trained alternately with opposing objectives; training is notoriously unstable (mode collapse, where the generator produces few varieties; oscillation) and needs careful engineering
- Typical successors
- diffusion models replaced GANs for general image and video generation because they train more stably and cover the data's variety better; GANs persist where speed at generation matters and the domain is narrow
- Parties
- the researchers (Ian Goodfellow's 2014 paper), the companies that productised face and video synthesis, the fraudsters who used them for identity attacks, and the detection teams
How a generative adversarial network works in practice
Start with a generator that produces noise and a discriminator that can barely tell noise from a photograph. Show the discriminator real faces and the generator's outputs; it learns to tell them apart; its gradients tell the generator which way to improve; the generator produces slightly better faces; the discriminator learns again. Thousands of rounds later the generator produces faces the discriminator cannot distinguish from real — and, in StyleGAN's case, that people cannot either, with controls over age, pose and expression in the generator's latent space. The same recipe swapped faces in video (deepfakes), upscaled images, and generated synthetic medical images for training where real data was scarce.
The fintech uses run both ways. Defensively: a fraud team training a model on rare fraud patterns can use a GAN to generate synthetic examples that augment the data, or synthetic transaction data that lets engineers develop and test without exposing real customer records. Offensively, from the fraudster's side: GAN-generated faces on synthetic identities, face-swapped videos to defeat video KYC, and — the general lesson — an adversary that adapts to whatever the detector learns. That adversarial framing is now standard in fraud modelling: the model is the discriminator, the fraudsters are the generator, and the model that is not retrained loses.
For AI M&A the GAN era's companies were the 2018–2021 synthetic-media and synthetic-data startups, many since pivoted to diffusion; for fintech the legacy is the identity-verification vendors' detection stacks and the synthetic-data vendors' products.
Example
A fintech's fraud team has 4,000 confirmed cases of a new account-takeover pattern among 40 million transactions — too few to train a robust model. It uses a GAN trained on the confirmed cases, conditioned on their features, to generate 60,000 synthetic examples that preserve the pattern's statistics without copying any real customer's record; the fraud model trained on the augmented data catches 82% of the next month's cases against 51% for the model trained on the 4,000 alone. The data-science team documents the synthetic data's provenance for the model-risk committee, which asks whether the GAN could have generated patterns that do not occur; the validation on held-out real cases is the answer.
Common mistakes
- Using GAN-augmented training data without validating on real held-out cases. - Treating GAN training as routine; instability is the norm and the results need inspection. - Reading diffusion's dominance as GANs' irrelevance for narrow, fast tasks. - Identity-verification flows built before synthetic faces existed and never updated. - Fraud models trained once and never retrained against an adversary that adapts by design.