Private Evolution: Generating Differentially Private Synthetic Data without Model Training
Private Evolution (PE) is a framework designed to generate differentially private (DP) synthetic data without requiring model training. Developed by Microsoft Research, PE leverages blackbox inference APIs of foundation models such as Stable Diffusion or GPT, as well as simpler data generation tools like simulators or random number generators. The framework ensures that synthetic data adheres to strict privacy guarantees, making it suitable for applications where sensitive data must remain protected.
The core innovation of PE lies in its ability to generate synthetic data while maintaining differential privacy. Unlike traditional methods that rely on training machine learning models, PE uses existing foundation model APIs or simulators to produce data that statistically resembles real datasets but does not expose underlying private information. This approach eliminates the need for extensive computational resources typically associated with model training, including GPU dependencies for certain implementations.
PE has been validated through multiple research papers published in leading conferences, including ICLR 2024, ICML 2024 Spotlight, and ICML 2026. These papers cover different data modalities, including images, text, tabular data, and simulations. The framework’s modular design allows it to be adapted for various use cases while preserving privacy guarantees.
Its flexibility and efficiency make it a practical solution for industries handling sensitive data, such as healthcare, finance, and government.
How It Works
Private Evolution operates by applying differential privacy techniques to the process of synthetic data generation. The framework does not require access to the internal parameters of foundation models; instead, it interacts with these models through their inference APIs. This blackbox approach ensures compatibility with a wide range of pre-trained models without the need for retraining or fine-tuning.
PE applies differential privacy mechanisms to perturb or modify this data in a controlled manner, ensuring that no individual record can be linked back to its original source.
For image generation, PE uses foundation models like Stable Diffusion to produce synthetic images. The framework incorporates privacy-preserving techniques such as differential privacy-aware sampling and noise injection. Similarly, for text generation, PE employs GPT-based models to create synthetic text while ensuring that the output does not reveal details about the input data. In the case of tabular data, PE relies on simulators or statistical methods to generate synthetic records without exposing private attributes.
The modular architecture of PE allows users to extend its functionality by integrating additional data generation tools or customizing privacy parameters to meet specific requirements.
Use Cases
Private Evolution is designed to address the growing demand for synthetic data in applications where privacy is a critical concern. One primary use case is in healthcare, where synthetic patient data can be used for training machine learning models without compromising patient confidentiality.
Banks and financial institutions can use this data to develop and test algorithms without exposing real customer information.
For research purposes, PE provides a tool for data scientists and AI researchers to explore new algorithms without relying on proprietary or restricted datasets. The framework’s ability to generate synthetic data across multiple modalities—images, text, and tabular data—makes it versatile for testing and validation in diverse domains.
Another emerging use case is in the development of AI systems that require large-scale training data but cannot ethically or legally use real-world datasets. By generating differentially private synthetic data, organizations can mitigate risks associated with data leakage while still achieving their AI development goals. PE’s compatibility with foundation model APIs also allows it to integrate seamlessly with existing AI pipelines.
Modality Support
PE supports the generation of synthetic data across four primary modalities: images, text, tabular data, and simulated data. For images, PE uses foundation models like Stable Diffusion to produce high-quality synthetic images while applying differential privacy techniques to obscure sensitive features. The Cat Cookie and Cat Doudou datasets, created as part of the initial research, demonstrate the framework’s capability in generating visually realistic synthetic images.
For text generation, PE integrates with GPT-based models to create synthetic text data. The framework ensures that the generated text maintains coherence and relevance while adhering to privacy constraints.
Tabular data generation is a significant advancement in PE, as it enables the creation of synthetic records for structured datasets without the need for GPU acceleration. The tabular data generation capability was released in July 2026 as part of the ICML 2026 paper, expanding PE’s applicability to domains such as finance, healthcare, and logistics.
Differential Privacy Guarantees
A defining feature of PE is its rigorous adherence to differential privacy principles. The framework ensures that the synthetic data generated does not reveal sensitive information about the original dataset. Users can adjust these parameters to balance between data utility and privacy protection, depending on their specific requirements.
No Model Training Required
Unlike traditional synthetic data generation methods that rely on training machine learning models, PE eliminates the need for this computationally intensive step. By leveraging existing foundation model APIs or simulators, PE achieves high-quality synthetic data generation without requiring access to model weights or training data. This reduces the barrier to entry for organizations that lack the resources or expertise to train custom models.
The absence of model training also simplifies the deployment of PE, as it can be integrated into existing workflows with minimal modifications. Organizations can generate synthetic data on-demand, scaling production as needed without the overhead of maintaining training infrastructure.
Modular and Extensible Architecture
PE’s architecture is designed to be modular, allowing users to extend its functionality by integrating additional data generation tools or customizing privacy mechanisms.
For example, users can replace the default foundation model APIs with alternative models or simulators, provided they meet the framework’s compatibility requirements. Similarly, privacy parameters can be fine-tuned to align with specific regulatory or organizational standards. This extensibility ensures that PE remains relevant as new data modalities and privacy challenges emerge.
The framework’s documentation and codebase are structured to facilitate customization, with detailed examples and tutorials available for each modality. Users can refer to the original research papers or the deprecated branches of the repository for historical implementations, ensuring transparency and reproducibility.
CPU-Compatible Implementations
One of the standout capabilities of PE is its ability to generate synthetic data on CPUs, eliminating the need for GPU acceleration in certain modalities. The tabular data generation feature, for instance, runs entirely on CPUs, making it accessible to a broader range of users without compromising performance or privacy guarantees.
Adoption and Integration
PE has been adopted by major technology companies, including Microsoft and Apple, demonstrating its practicality and scalability in real-world settings. The framework’s integration into production systems highlights its ability to meet industry standards for privacy and data utility. Organizations can leverage PE to enhance their data strategies while mitigating risks associated with sensitive information exposure.
The framework’s compatibility with existing foundation model APIs ensures smooth integration into AI pipelines. Users can generate synthetic data for training, testing, or validation purposes without disrupting their current workflows. PE’s modular design also allows for incremental adoption, enabling organizations to start with one modality and expand to others as needed.
Responsible Use and Considerations
While PE provides robust privacy guarantees, users must consider potential risks and limitations when deploying synthetic data in downstream applications. Synthetic data may not perfectly replicate the statistical properties of real-world datasets, which could impact the performance of machine learning models trained on such data.
Additionally, foundation model APIs and simulators used by PE may introduce biases or inaccuracies, which could propagate into the synthetic data.
PE’s documentation emphasizes the importance of responsible use, encouraging users to conduct thorough testing and validation before deploying synthetic data in production environments. By adhering to best practices, users can maximize the benefits of PE while minimizing potential harms.