Generative Adversarial Network (GAN)
Definition
A machine learning architecture consisting of two networks trained simultaneously: a generator that produces synthetic images and a discriminator that classifies images as real or synthetic. Adversarial training drives both networks toward equilibrium, producing increasingly convincing output.
- Components
- Generator network and discriminator network
- Training method
- Adversarial, simultaneous
- Introduced
- Ian Goodfellow et al., 2014
- Forensic relevance
- Common source of synthetic deepfake images
Common questions
How do the generator and discriminator interact during training?+
The generator produces synthetic images while the discriminator tries to tell them apart from real ones, and the resulting competition pushes the generator toward increasingly realistic output as it learns to fool the discriminator.
What forensic artifacts can betray a GAN-generated image?+
Common tells include unnatural symmetry, inconsistent lighting or reflections, blending boundaries around swapped regions, and characteristic frequency-domain patterns left by the network's upsampling layers.
Are GANs the only architecture used to generate synthetic faces?+
No, diffusion models are a major alternative that generate images through iterative denoising rather than adversarial competition, and they leave different forensic artifacts than GANs do.
Related terms
- Autoencoder
- A neural network that compresses an input into a compact latent representation (encoder) then reconstructs it (decoder). Face-swap pipelines train autoencoders with...
- Blending Mask
- In face-swap pipelines, a pixel-level mask that defines the face region to be composited onto the target frame. Imperfect masks leave boundary...
- Checkerboard Artifact
- A grid-like pattern visible in the Fourier power spectrum of GAN outputs, caused by transposed convolution or bilinear-upsampling operations used to increase...
- Diffusion Model
- A generative neural network architecture (Ho et al., 2020; Stable Diffusion, Rombach et al., 2022) that learns to reverse a noise-addition process...
- Latent Space
- The compressed, lower-dimensional representation of data learned by a neural network's internal layers. Generative models sample from or navigate this space to...