Week 14 · Unit 2, Lesson 4 Artificial Intelligence and Deepfake · 3 lessons
2
Two networks that compete: generator and discriminator
ICT Prep 3 Term 1, Unit 2, Lesson 4: pp. 88–95
GoalExplain how two networks, a generator and a discriminator, compete and improve each other.
New words
GAN · generative adversarial network: two networks that competegenerator · the network that makes fake examplesdiscriminator · the network that decides whether an example is real or faketraining round · one turn of making, judging and improving
Explanation
One technique for making deepfakes is a GAN. It uses two neural networks that compete, like a forger and an inspector. The generator is the forger: it makes a fake, for example a picture of a face. The discriminator is the inspector: it looks at real examples and the fake and says real or fake. Then the inspector’s answer goes back to the forger, who tries to make a better fake. In the next round the inspector has to look harder. After many rounds the fakes can become very close to real, and the inspector is sometimes wrong. A GAN is one technique among several, which is also why we should not trust any single trick to spot a fake.
Worked examples
A forger copies a painting and an inspector says which ones are fake. At first the copies are clumsy, so the inspector is sure. As the forger learns from the answers, the copies improve, and the inspector must look more closely. Both get better together.
Round (invented)
How the fake looks
The inspector says
1
clumsy, easy to spot
fake, with confidence
20
better, still has flaws
fake, less sure
200
very close to real
sometimes wrong
The table gives invented numbers for three rounds of a GAN. Early on the inspector is sure; later the fake is so close that the inspector is sometimes wrong. This is why a fake that looks perfect to our eyes is possible.
Class activity
Draw the cycle of a GAN on paper with the forger, the inspector and the arrows. Write one sentence about what changes in each round. Then explain, in your words, why a GAN does not mean that every fake can be found by one trick.
Practice
1
Which network is the forger: the generator or the discriminator?
The generator, because it makes the fakes.
2
What does the discriminator decide?
Whether an example is real or fake.
3
What does the generator receive after each round?
The discriminator’s answer, so it can make a better fake.
