NEWS

RGAN 0.2.0

New dp_gan_trainer() function for training GANs with formal differential privacy guarantees using DP-SGD (Differentially Private Stochastic Gradient Descent)
Privacy mechanisms include:
- Poisson subsampling for privacy amplification
- Per-sample gradient clipping to bound sensitivity
- Calibrated Gaussian noise addition
- Rényi Differential Privacy (RDP) accounting for tight composition
New secure_rng parameter to choose between cryptographically secure RNG (OpenDP)

for production use or fast torch RNG for development/testing
Privacy budget pre-calculation with compute_max_steps() for efficient training
New calibrate_noise_multiplier() to find optimal noise for target epsilon
New vignette: "Training GANs with Differential Privacy"

New apply_post_gan_boosting() wrapper for easy post-processing of GAN samples
New compute_discriminator_scores() for evaluating samples across checkpoints
Checkpoint support in both gan_trainer() and dp_gan_trainer():
- checkpoint_epochs parameter to save model states at intervals
- checkpoint_path for disk-based storage of large training runs
Support for differentially private post-GAN boosting with exponential mechanism
End-to-end privacy guarantees when combining DP training with DP boosting
New vignette: "Post-GAN Boosting for Improved Synthetic Data Quality"

Comprehensive documentation for data_transformer class including:
- Detailed explanation of standard vs mode-specific (GMM) normalization
- Examples for mixed continuous/categorical data
- Integration examples with RGAN workflow
Added CTGAN reference (Xu et al., 2019) for mode-specific normalization

New track_loss parameter in gan_trainer() and dp_gan_trainer() to record training losses for analysis
Loss tracking available via trained_gan$losses$g_loss and $d_loss

OpenDP package now suggested (optional) for cryptographically secure DP training