Differential privacy (DP) has been used for nearly two decades to help protect sensitive individual information in datasets used for analysis. But applying DP to every analytical technique can be complex, burdensome, and error-prone.
A new approach described in the source text uses generative AI models like Gemini to create a single private synthetic version of the original dataset. The synthetic data is made from common data patterns and is designed to contain no unique details from any individual user.
The method uses a differentially private training algorithm, such as DP-SGD, to fine-tune the generative model on the original dataset. That allows the synthetic dataset to remain private while staying highly representative of the real data. Standard, non-private analytical techniques or modeling can then be run on the substitute dataset.
The source text says this approach is especially useful when organizations need high-volume, controlled datasets but do not have access to high-quality, representative data. It also notes that much of the published work on private synthetic data has focused on simple outputs such as short text passages or individual images, while modern multi-modal applications often need richer data.
To address that gap, the method introduced in the source text generates synthetic photo albums from rich, structured image-based datasets. The task is more complex than generating single images because it must preserve thematic coherence and character consistency across multiple photos in a sequence.
The method translates complex image data to text and back. The source text says the process, with rigorous DP guarantees enabled, preserves high-level semantic information and thematic coherence needed for analysis and modeling applications.
Source: research.google.
Companies can share verified announcements through Newz9’s international press release submission page.
