Generative Adversarial Network Architecture
- 27/06/2025 by Abdi Dharma
Generative Adversarial Network or commonly abbreviated as GAN is a deep learning architecture conceptualized in 2014 by Ian Goodfellow and several other researchers, and this technique is very popular, because it can produce various fake data using the original data available, thus creating the term Deepfake. Another application that can be produced from this GAN technique is on the face image page via the This Person Does Not Exist site, the face images produced from this site are face images created using the GAN technique, and over time, this technique has become more and more mature, because the resulting images are very realistic, and seem to be photographed from a real human face.
GAN DIAGRAM
Simply put, the GAN architecture can be described in the image above, and the main actor in this architecture is the Generator model, while the supporting actor is the Discriminator model. The way this technique works is actually very simple and can be illustrated in a story between a criminal and a police officer.
CRIMINAL AND POLICE
Assume the Generator model is a Criminal who has a Printing Machine and is tasked with printing Counterfeit Money, while the Discriminator model is a Policeman who has a Scanning Machine and is tasked with detecting the difference between Real Money and Counterfeit Money.
The Criminal will take a blank paper containing random data (Random input), and from this random paper, it will go into the Printing Machine (Generator) to produce Counterfeit Money (Fake Samples).
Along with this Counterfeit Money, Real Money (Real Samples) will be mixed together and handed over to the Police (Discriminator) to be analyzed by the Scanning Machine.
The Police themselves know which is Counterfeit Money and which is Real Money, because this money is labeled, and in this position, the police will detect and evaluate the results of their predictions and analysis (Loss).
If you are still not satisfied (Satisfy?) with the evaluation results (Loss), then the results will be given to the Criminal (Generator) and the Police (Discriminator) to analyze the evaluation results and adjust their respective machines (Fine Tuning) with the aim of improving the Criminal's ability to make Counterfeit Money that is similar to Real Money, and for the Police to be better at distinguishing Real Money and Counterfeit Money (Adversarial).
If you are satisfied (Satisfy?), then at the end of this cycle, the Criminal (Generator) will be the main actor in taking blank paper containing random data (Random Input) to produce Counterfeit Money (Generated Data) that is similar to Real Money to be circulated in the market.
So in this case, there are two models that are the benchmarks in the GAN technique, namely the discriminator model, where this model is tasked with making predictions and learning, and there is a generator model that is tasked with generating data, and these two models seem to compete with each other, and in mathematical game theory, the process of competing between these two models is also called a zero-sum game. Zero-sum game is a match between two parties, where the result of this match, the winner will get the same level of profit as the loss on the loser, so if the winner gets +3, then the loser will get -3, so the final result is actually 0, and that's why this game is called zero-sum, where if you add the results of the winner and the loser, it will still produce a value of zero.
And also in accordance with the meaning of this technique itself, where the word generative means that this technique produces something, which is continued with the word adversarial, which means hostile to each other, and closed with the word network or network, so if written in full, generative adversarial network, meaning a network that works in a hostile way to produce something.