Introduction
Training deep learning models with Differential Privacy (DP) has become the gold standard for protecting sensitive user data. However, this protection often comes with a significant "utility tax." The standard algorithm, DP-SGD, injects noise into gradients during training to ensure privacy, which frequently leads to a substantial drop in accuracy compared to non-private counterparts.
To mitigate this, researchers often utilize "warm-up" training on public datasets. A major hurdle arises when public data is out-of-distribution (OOD)—for instance, using general ImageNet photos to pre-train a model intended for specialized medical imaging. While having in-distribution (ID) public data (data that mirrors the private set) is ideal, it is often only available in tiny quantities. In such cases, models tend to overfit the small public sample, failing to generalize to the larger private dataset. This paper introduces a framework to solve the "limited public data" problem by amplifying a small public subset to guide private training effectively.
The WarmSyn Framework
The authors propose a three-stage pipeline called WarmSyn, designed to maximize the utility of a tiny sliver of public data (as little as 4% of the total dataset).
- Diffusion Synthesis: The process begins by training an Elucidating Diffusion Model (EDM) on the small public subset. This model is then used to generate a massive synthetic dataset (e.g., expanding 2,000 images into 50,000).
- Diversity Enhancement: To prevent the model from merely memorizing the small public set, the authors employ Discriminator Guidance (DG) and aggressive data augmentations like Cutout and Cutmix. This forces the diffusion model to explore the broader "data manifold," prioritizing diversity over perfect visual fidelity.
- Sharpness-Aware Warm-up: The model is pre-trained on this synthetic data using Sharpness-Aware Minimization (SAM). SAM identifies "flat" regions in the loss landscape, which are known to generalize better. This helps bridge the gap between synthetic public data and real private data.
- Private Fine-tuning: Finally, standard DP-SGD is performed on the actual private data, starting from the robust initialization provided by the warm-up phase.
Key Insights: Why It Works
Diversity Over Fidelity
One of the paper's most striking findings is that for classification tasks, synthetic images do not need to be high-fidelity replicas of the original data. Instead, having a diverse set of images—even if they appear "noisy"—is more beneficial for the model's ability to learn generalizable features. High discriminator guidance was essential in achieving this variety.
Bridging the Generalization Gap
There is a distinct performance gap between public synthetic data and private real data. By using SAM to find flatter minima during the public warm-up, the model becomes significantly more resilient to the noise added during the subsequent private training phase.
As shown in the analysis of training dynamics, the WarmSyn approach leads to smaller and more stable gradient norms during the private phase. Since DP-SGD performance is highly sensitive to gradient clipping, this "natural" reduction in norm size allows the model to learn more effectively from private samples without losing information to aggressive clipping.
Experimental Results
The researchers evaluated WarmSyn across several benchmarks, consistently achieving state-of-the-art (SOTA) results:
- CIFAR-10: Under a strict privacy budget of $\epsilon=2$, the method achieved 85.48% accuracy, a massive leap from previous SOTA methods which hovered around 75.1%.
- CIFAR-100: On this more complex dataset, they reached 48.61% accuracy at $\epsilon=2$ without any external pre-trained models—a significant improvement over the ~18% baseline.
- Real-World Datasets: On EuroSAT (satellite imagery), accuracy improved from ~84% to over 90% at $\epsilon=2$. Similar gains were observed on PathMNIST (biomedical) and FairFace.
Conclusion
WarmSyn demonstrates that "amplifying" a tiny amount of in-distribution public data through diffusion models is more effective than traditional optimization-heavy DP methods. By focusing on data diversity and using sharpness-aware optimization to find robust minima, the framework significantly reduces the "utility tax" of differential privacy, bringing private models closer to the performance of their non-private counterparts.
Summary updated