Rényi differential privacy has become a popular framework for quantifying privacy loss, but its core calculation—Rényi divergence—is often intractable except in Gaussian cases. This limits its use for broader classes of mechanisms.
The paper introduces "task-aware kernel Stein certificates" to certify Rényi DP. These certificates appear to sidestep direct divergence computation, instead using kernel-based statistical tools tailored to the specific task or query.
By making certification more tractable, the approach could extend Rényi DP to settings where it was previously impractical. The abstract does not provide experimental details, so the practical gains remain to be seen.