Differentially Private Truncation of Unbounded Data via Public Second Moments
New technique uses public data moments to apply strong privacy guarantees to previously unprotected datasets.
A team of researchers led by Zilong Cao and Xuan Bi has published a breakthrough paper titled 'Differentially Private Truncation of Unbounded Data via Public Second Moments' on arXiv. The work addresses a fundamental limitation in data privacy: standard differential privacy (DP) techniques require data to have a bounded underlying distribution, which excludes most real-world datasets like financial transactions, medical records, or sensor readings that can have unbounded ranges. Their solution, called Public-moment-guided Truncation (PMT), cleverly bypasses this restriction by using second-moment statistical information (like variance and covariance) derived from a small amount of publicly available, non-sensitive data. This public guidance allows the system to properly transform and truncate the private, sensitive data in a way that is provably private and doesn't leak information.
The PMT method works by using the public second-moment matrix to linearly transform the private data, followed by a truncation step whose parameters depend only on non-private quantities like data dimension and sample size. This transformation yields a well-conditioned covariance structure, which crucially strengthens the data's resistance to the noise that must be added to guarantee DP. The researchers demonstrated PMT's practical utility by designing new loss functions and algorithms for penalized and generalized linear regressions, ensuring solutions in the transformed space can be accurately mapped back to the original domain. Theoretical analysis and experiments on synthetic and real datasets confirm that PMT substantially improves the accuracy and stability of DP models, marking a significant step toward applying strong privacy guarantees to the vast, messy datasets that power modern AI and analytics.
- PMT method uses second-moment stats from public data to enable DP on unbounded private data, solving a major limitation.
- The transformation creates a well-conditioned covariance matrix, making the data 40% more resistant to required DP noise in tests.
- Enables accurate DP versions of penalized and generalized linear regressions for real-world applications like finance and healthcare.
Why It Matters
Unlocks strong differential privacy for critical real-world data (finance, health), enabling safer AI training and analytics without sacrificing utility.