3rd IEEE Conference on Secure and Trustworthy Machine Learning (SaTML2025): Lorenz Wolf

I recently had the pleasure of presenting our work “Private Selection with Heterogeneous Sensitivities” at the 3rd IEEE Conference on Secure and Trustworthy Machine Learning (SaTML2025) in Copenhagen. This work was presented in the form of a poster as well as a talk at the main conference.

Our paper tackles a key challenge in differential privacy: how to select the best option (e.g., a model or hypothesis) when different candidates have unequal sensitivities to the data. This kind of heterogeneity is common in real-world tasks like recommendation systems or hyperparameter tuning—but standard private algorithms like Report Noisy Max (RNM) don’t account for it.

First we show that adding heterogeneous noise despite introducing less randomness is not always beneficial in terms of utility and that existing mechanisms can behave drastically differently depending on the distribution of scores and sensitivities. We then introduce a new mechanism, mGEM, which performs well when high-scoring candidates are also more sensitive. We also propose a correlation-based heuristic to guide mechanism choice, using the relationship between scores and sensitivities. Finally, our combined approach that adaptively and privately selects between GEM and mGEM based on this heuristic performs well in polarized settings, though creating a trade-off between algorithm choice and the performance of the chosen algorithm.  On datasets like Netflix, MovieLens, and Amazon Books, our mGEM outperforms existing methods.

During the Q&A two interesting questions came up: Why do scores and sensitivities tend to be positively correlated in practice? There was also interest in the details of our adaptive mechanism selection. This is an area we’re actively refining and trying to further improve performance!

I am grateful for my collaborators and the SaTML community for such an inspiring event.