UCR pioneers way to remove private data from AI models

Researchers from UC Riverside have developed a "source-free certified unlearning" method that allows AI models to erase specific private or copyrighted data without requiring access to the original training dataset. By using a statistically similar surrogate dataset and calibrated noise, the system effectively compels the model to forget targeted information while maintaining its overall functionality. This breakthrough solves a critical limitation in current AI governance, as it enables data removal even when the initial training records are lost or inaccessible, which is a common real-world scenario. The urgency for such technology is driven by stringent privacy regulations like the GDPR and ongoing legal battles over copyright infringement in generative AI. Traditional methods of removing data typically require costly and energy-intensive full retraining from scratch. The new approach offers a significantly more efficient alternative, providing privacy guarantees comparable to full retraining but with far less computational power. This efficiency makes it feasible for developers to comply with legal obligations and ethical standards without prohibitive resource costs. This development is highly relevant to the open data community because it establishes a practical mechanism for ensuring data sovereignty and ethical AI usage. It empowers individuals and organizations to demand the deletion of their contributions from public models, fostering trust in open-source and public AI systems. By providing a provable way to remove sensitive information, the technique supports the responsible sharing of AI technology, ensuring that open data ecosystems respect individual rights and legal boundaries while remaining scalable for future complex models.

Source: pressreleasepoint.com
Published on 2025-08-30

UCR pioneers way to remove private data from AI models

Related news