Generative Adversarial Network

Definition

A generative adversarial network (GAN) is a deep learning framework introduced by Ian Goodfellow and colleagues in 2014. It consists of two neural networks trained simultaneously in a competitive game: a generator network that produces synthetic samples (images, audio, text, video) and a discriminator network that attempts to classify whether a given sample is real or generated. Through iterative training, the generator improves at producing realistic outputs while the discriminator improves at detecting fakes, until the generator produces outputs that the discriminator cannot reliably distinguish from real data. GANs were the first architecture to produce photo-realistic synthetic images at scale and remain widely used despite the emergence of diffusion models.

How It Relates to AI Threats

GANs are a core enabling technology within the Information Integrity Threats domain. The deepfake phenomenon — face-swapping in video, AI-generated portrait images, synthetic voice recordings — was originally powered primarily by GAN architectures. While diffusion models have since become the dominant approach for high-quality image generation, GANs remain relevant for real-time applications (live video face-swapping), specialised use cases (super-resolution, image inpainting), and many freely available deepfake tools that still rely on GAN-based pipelines. Understanding GAN architecture is essential for comprehending how synthetic media threats emerged and how detection systems work.

Why It Occurs

• The adversarial training dynamic between generator and discriminator produces increasingly realistic synthetic outputs
• GAN architectures are computationally efficient for real-time generation compared to diffusion models
• Pre-trained GAN models and open-source implementations are widely available, lowering the barrier to generating synthetic media
• The same adversarial principle that makes GANs effective for generation also makes GAN-generated content difficult to detect
• Rapid progress in GAN research continuously improves output quality while reducing the data and computation required

Real-World Context

GANs powered the first wave of deepfake videos that gained public attention in 2017-2018. StyleGAN (NVIDIA, 2019) demonstrated the generation of photo-realistic human faces that fooled human observers. GAN-based tools have been used in documented financial fraud incidents, non-consensual intimate imagery generation, and election interference campaigns. While diffusion models now produce higher-quality outputs for many tasks, GAN-based deepfake tools remain prevalent in freely available software and continue to be used in threat incidents documented in the TopAIThreats taxonomy.