Deep generative models, particularly Generative Adversarial Networks (GANs), are highly effective at synthesizing realistic data. However, they are notorious for memorizing training samples, posing severe privacy risks. While Differential Privacy (DP) offers a rigorous mathematical framework to guarantee individual privacy, integrating DP into adversarial training introduces a fundamental conflict.
Standard private optimization (such as DP-SGD) requires clipping gradients and injecting stochastic noise to mask the influence of any single data point. When this noise enters the delicate, zero-sum game of adversarial training, it destabilizes the optimization landscape, frequently causing catastrophic divergence, mode collapse, and poor sample utility.
The Core Conflict: Minimax Games Meet Privacy Noise
To understand why DP destabilizes adversarial training, we must look at the feedback loop between the generator and the discriminator. In a standard GAN, the generator relies on precise, fine-grained gradient feedback from the discriminator to learn the underlying data distribution.
As noted by Marcel Neunhoeffer et al. in Private Post-GAN Boosting, injecting privacy noise into these gradient updates obscures the subtle signals the generator needs to learn complex structures. Instead of converging to a stable equilibrium, the training trajectory becomes highly volatile. This often leaves researchers with a final model state that is significantly worse or more biased than intermediate states captured earlier in training.
Furthermore, Dingfan Chen et al. demonstrate in A Unified View of Differentially Private Deep Generative Modeling that applying privacy constraints to a dynamic component—like a discriminator that is constantly changing its own parameters—compounds convergence failures. The noise turns the discriminator's feedback into a moving, heavily blurred target.
The Batch Size Dilemma
The instability is further exacerbated by a structural mismatch in hyperparameter requirements, specifically regarding batch sizes.
Standard GAN Training: Small Batch Sizes ---> Stable Discriminator Dynamics
Private Training (DP): Large Batch Sizes ---> Mitigates Impact of Privacy Noise
As explained by Chih-Hsun Lin et al. in DPAF: Image Synthesis via Differentially Private Aggregation in Forward Phase, DP mechanisms require large batch sizes to average out the injected noise and maintain a reasonable privacy budget. However, adversarial frameworks typically require smaller batch sizes to maintain stable discriminator performance. Forcing a large batch size onto the discriminator to satisfy DP constraints degrades the training dynamics, resulting in blurry, "blob-like" synthetic outputs.
Additionally, when the underlying data distributions are complex or disjoint, standard adversarial metrics like KL-divergence yield weak training signals under noise, failing to drive the model toward total uncertainty or random guessing (Sheikh Shams Azam et al. in Can we Generalize and Distribute Private Representation Learning?).
Bypassing the Discriminator with Stable Objectives
To resolve this instability, a major branch of research abandons the traditional minimax game entirely, replacing the unstable discriminator with stable, non-adversarial distance metrics.
Convex-Regularized Distance Metrics
Instead of training a neural network discriminator to find differences between distributions, Tianshi Cao et al. introduce Don't Generate Me: Training Differentially Private Generative Models with Sinkhorn Divergence. By optimizing a "semi-debiased" Optimal Transport objective (Sinkhorn divergence), they replace the competitive game with a stable, convex-regularized objective. This allows for reliable convergence across a wide range of hyperparameters without the risk of catastrophic divergence.
One-Shot Privatization
Taking the elimination of iterative adversarial training a step further, Frederik Harder et al. propose DP-MERF: Differentially Private Mean Embeddings with RandomFeatures for Practical Privacy-preserving Data Generation. Instead of adding noise at every single training step, DP-MERF privatizes the target data distribution in a single, "one-shot" step using random feature approximations of kernel mean embeddings. The generative model then simply minimizes the distance to this static, pre-privatized target, completely bypassing the dynamic instability of private GANs.
Energy-Guided Sampling
Similarly, Chia-Mu Yu et al. develop DPGEN: Differentially Private Generative Energy-Guided Network for Natural Image Synthesis. They break the low-resolution barrier of private image synthesis by utilizing a privatized energy-guided network to direct Langevin MCMC sampling, avoiding the equilibrium failures inherent to standard private GAN frameworks.
Decoupling Privacy from Gradient Updates
Another powerful intuition for stabilizing private generative models is to decouple the privacy mechanism from the backpropagation loop entirely. Rather than sanitizing the gradients during backpropagation (which directly corrupts the optimization step), researchers inject noise into alternative phases of the pipeline.
- Forward Phase Aggregation: The authors of DPAF shift the noise injection step from the training gradients to the model's forward pass. This preserves the integrity of the gradient feedback loop, significantly reducing information loss and allowing the model to generate high-fidelity structural details.
- Slicing Mechanisms: Kristjan Greenewald et al. introduce Privacy without Noisy Gradients: Slicing Mechanism for Generative Model Training. This method injects noise into random low-dimensional projections (slices) of the data rather than the gradients. Because the optimization loop itself remains noise-free, models can be trained for unlimited epochs and hyperparameters can be tuned freely without exhausting the privacy budget.
Alternative Architectures and Public Pre-training
When neural network optimization under DP remains too volatile, researchers either constrain the architecture or bootstrap the model using external data.
Algorithmic Robustness via Probabilistic Circuits
Hyperparameter tuning under DP is uniquely challenging because selecting the best model from a pool of candidates consumes a portion of the finite privacy budget. To address this, Xenia Heilmann et al. introduce Differentially Private Sum-Product Networks. By abandoning neural networks in favor of Sum-Product Networks (probabilistic circuits), they achieve high algorithmic robustness and low variance across runs, ensuring stable convergence without wasting the privacy budget on hyperparameter search.
Public-to-Private Transfer Learning
Rather than attempting to learn complex data distributions entirely from scratch under heavy noise, Ruihan Wu et al. demonstrate in Large-Scale Public Data Improves Differentially Private Image Generation Quality that a generative model can be pre-trained on generic, large-scale public data (like ImageNet) and subsequently adapted to a private domain. This transfer learning approach bypasses the destructive initial phases of private adversarial training, yielding state-of-the-art image utility even under strict privacy guarantees.
Key Foundational Work
Private Post-GAN Boosting (2021) — contributed 2 concept elements to this direction.
Differentially Private Generative Adversarial Networks with Model Inversion (2021) — contributed 2 concept elements to this direction.
Don't Generate Me: Training Differentially Private Generative Models with Sinkhorn Divergence (2021) — contributed 2 concept elements to this direction.