SECAI Core · phase 2 of 15 · noun
Generative adversarial network
A type of artificial intelligence model in which two neural networks compete against each other to generate realistic synthetic data, such as images or text.
The Explain card
- Plain English
- A generative adversarial network (GAN) pits two neural networks against each other: a generator makes synthetic data and a discriminator tries to spot the fakes. Each improves by trying to beat the other.
- Example
- A GAN trained on employee photos produces realistic faces for a fake social media profile used in a social engineering campaign against the company.
- Why it matters
- GANs helped create the deepfake era. Understanding the generator-versus-discriminator loop also explains why detection is an arms race: every better detector trains a better forger.
- Hook
- Forger and art expert locked in a room until the forger wins.
Word knowledge
How the word is built, where it came from, and what it sits beside in memory.
In a sentence
The identity-proofing bench trains a generative adversarial network on passport portraits, then the onboarding camera scores each synthetic face as live or spoofed.
Why these words
- generative Latin generativus, of begetting, from generare, to beget marks the named system as one that produces new samples rather than sorting existing ones
- adversarial Latin adversarius, opponent, from adversus, turned against marks the training as a contest between two models rather than a single shared loss
- network English net plus work, a fabric of knots supplies the head noun, the computational object the other two words describe
Where it came from
- Origin
- English computing compound by Ian Goodfellow and colleagues, from Latin generativus (of begetting, from generare, to beget), Latin adversarius (opponent, from adversus, turned against), and English network (a fabric of knots)
- Entered the language
- 2014
- What changed
- Generative had meant able to beget since the 1400s, then able to produce sentences in Chomsky's grammar, then able to draw samples in statistics. Adversarial had meant of an opponent in court, then of a training setup that attacks its own model. Network had meant a fabric of threads, then any lattice of connections. Goodfellow's 2014 paper joined the three as Generative Adversarial Nets, English later lengthened Nets to network, and the phrase became the ordinary label for that two-player training setup.
How it is spelled
- Pattern
- verbs in -ate form adjectives in -ative
- Pattern
- -ary nouns sit beside -arial adjectives
- Pattern
- Latin vers keeps s, never c
- Pattern
- network is one closed compound, not net-work
- Breaks the pattern
- the noun phrase stays three words; hyphens appear only when the stack modifies another noun
- Breaks the pattern
- the 2014 paper wrote Nets; English later lengthened it to network in the standard phrase
Spelled like
- relative
- secretarial
- reverse
- internet
Broken into chunks
-
gen
- generate
- genesis
- progenitor
-
ative
- relative
- narrative
- innovative
-
advers
- adversary
- adverse
- adversity
-
ial
- commercial
- official
- industrial
-
net
- internet
- netting
- dragnet
-
work
- homework
- workshop
- artwork
What it sits beside
Same subject
Same shape
- convolutional neural network
- recurrent neural network
- graph neural network
Neural network architectures
Generative family
- generative AI
- generative model
- variational autoencoder
Adversarial family
- adversarial example
- adversarial training
- adversarial attack
Where it sits in the deck
Phase 2: How Models Are Built: Architectures and Learning Mechanics
With the paradigms named, zoom in on the architectural building blocks that turn data into trained artifacts.