Generative Adversarial Network (GAN)

A GAN consists of two competing networks locked in a minimax game. The generator creates synthetic samples from random noise, while the discriminator attempts to distinguish real data from fakes. This adversarial pressure forces the generator to produce outputs statistically indistinguishable from the training distribution. The training objective is formalized as:

• Generator loss: maximize the probability of the discriminator making a mistake
• Discriminator loss: maximize the probability of correctly labeling real and generated samples
• Nash equilibrium: the theoretical point where generated data perfectly matches the real distribution

The generator maps points from a latent space—typically a low-dimensional Gaussian distribution—to high-dimensional data outputs. This learned mapping creates a continuous manifold where:

• Interpolation between latent points produces semantically meaningful transitions
• Vector arithmetic in latent space enables feature manipulation (e.g., adding a 'smile' vector)
• Random sampling enables infinite synthetic data generation from a compact representation
• The smoothness of the latent space directly impacts the diversity and quality of generated samples

A common failure mode where the generator learns to produce only a limited variety of outputs that reliably fool the discriminator, ignoring large portions of the target distribution. Mitigation strategies include:

• Minibatch discrimination: allowing the discriminator to compare samples within a batch
• Unrolled GANs: using discriminator foresight to prevent generator exploitation
• Wasserstein loss: providing smoother gradients that discourage mode dropping
• Packing multiple samples into discriminator inputs to enforce diversity assessment

GAN training is notoriously brittle due to the non-stationary optimization landscape. As the generator improves, the discriminator's task shifts, creating a moving target. Key stabilization techniques include:

• Spectral normalization: constraining discriminator Lipschitz constant for stable gradients
• Two time-scale update rule (TTUR): using different learning rates for each network
• Gradient penalty: enforcing smooth discriminator behavior through regularization
• Progressive growing: gradually increasing resolution during training to build stable foundations before adding complexity

Conditional GANs (cGANs) extend the architecture by feeding auxiliary information—such as class labels, text descriptions, or images—to both generator and discriminator. This enables:

• Class-conditional synthesis: generating specific disease phenotypes from diagnostic labels
• Image-to-image translation: converting MRI to CT scans via paired supervision
• Text-to-image generation: creating medical illustrations from radiology reports
• The auxiliary input constrains the generator's output space, dramatically improving sample relevance for downstream clinical tasks

Assessing GAN output quality requires specialized metrics beyond simple loss curves. Standard evaluation approaches include:

• Inception Score (IS): measures both image quality and diversity using a pretrained classifier
• Fréchet Inception Distance (FID): compares feature distributions of real and generated samples—lower scores indicate higher fidelity
• Precision and Recall for Distributions: separately quantifies sample quality and coverage of the real data manifold
• Domain-specific validation: for clinical data, downstream task performance (e.g., diagnostic accuracy) serves as the ultimate utility metric

A generative model that learns a compressed latent representation of input data and reconstructs new samples by sampling from that learned distribution. Unlike GANs, which use an adversarial game, VAEs are based on variational inference and optimize a reconstruction loss plus a regularization term. They are often preferred for synthetic EHR generation when stable, smooth latent spaces are required, though their outputs can be blurrier than GAN-generated samples.

A mathematical framework that injects calibrated noise into data or model updates to provide a provable guarantee that individual patient records cannot be inferred. When applied to GAN training, it bounds the privacy loss (ε) an adversary can incur. This is essential for Federated GAN architectures, ensuring that the generator does not memorize and regurgitate specific training samples from any single hospital's dataset.

A decentralized architecture where the discriminator and generator are trained across multiple institutions without sharing underlying patient data. Each site trains a local discriminator, and only model updates—not raw images or records—are aggregated. This allows multi-site collaboration to generate high-fidelity synthetic medical images or tabular data while maintaining strict HIPAA and GDPR compliance.

An adversarial technique that determines whether a specific patient's record was used in training a model. For GANs, attackers analyze the discriminator's confidence or the generator's output distribution to identify overfitting artifacts. Defenses include differential privacy, dropout, and limiting the number of training epochs to prevent the generator from memorizing rare training examples.

A quantitative measure of how well a synthetic dataset preserves the statistical relationships and predictive performance of the original real-world data. Key evaluation paradigms include:

• Train-Synthetic-Test-Real (TSTR): Train on fake data, test on real data
• Train-Real-Test-Synthetic (TRTS): Train on real data, test on fake data
• Column-wise distribution comparison using Jensen-Shannon divergence

A conditional generative adversarial network specifically designed to model tabular data with mixed discrete and continuous columns. It uses mode-specific normalization to handle non-Gaussian distributions and a training-by-sampling strategy to address class imbalance. CTGAN is a core component of the Synthetic Data Vault (SDV) ecosystem and is widely used for generating realistic synthetic EHR tables.