I am excited to share that our paper, “Optimizing Canaries for Privacy Auditing with Metagradient Descent,” has been accepted at ICLR 2026!

You can find the full paper on arXiv here: arXiv:2507.15836

Our main contribution is a method for optimizing the “canary” examples used by auditors to detect privacy leakage, leveraging recent advances in metagradient optimization. Our empirical evaluation shows that these optimized canaries can improve empirical lower bounds for image classification models by over 2x in certain instances.

Our method utilizes the REPLAY algorithm to compute metagradients efficiently and at scale. We demonstrate that canaries optimized on a small ResNet-9 architecture remain highly effective even when transferred to audit much larger models like Wide ResNets trained with DP-SGD.

Stay tuned for the full camera-ready version and code release!