Mohammad Yaghini

The privacy-fairness tension is considerably weaker than previously believed. It appears largely to be an artifact of algorithm design. 🌐 cleverhans.io/dp-raco 📄 iclr.cc/virtual/2026/p… 💻 github.com/cleverhans-lab…

Unlike regularization-based methods, RaCO-DP enforces bounds directly by construction: no hyperparameter tuning, no extra privacy cost. Also scales to many simultaneous constraints without degradation, and runs three orders of magnitude faster than the prior SOTA, DP-FERMI.

Can ML models be trained under fairness constraints with formal DP guarantees, without sacrificing utility? New paper at ICLR 2026: RaCO-DP Poster: Pavilion 4 (later today!) 🧵 (w/ @tudorcebere, Michael Menart, Aurélien Bellet, @NicolasPapernot )

Curious about novel risks that AI introduces for patients and healthcare providers? We're hosting an #AIUK Fringe event to explore Privacy and Fairness in AI for Health – Register now! private-fair-ai.github.io Date: 27 March, 10:00-16:30 GMT Location: @turinginst, London 1/6

@arkrause Congratulations Andreas!

Thank you so much for this tremendous honor! Of course this award also recognizes the contributions of my amazing students, colleagues and collaborators whom I’ve been extremely fortunate to work with!

@HodaHeidari @mldcmu @isrcmu Congratulations Hoda.

I am thrilled to announce that I will be joining the School of Computer Science at Carnegie Mellon University as an assistant professor with joint appointments in the @mldcmu and @isrcmu!

@hiddenmarkov @Aaroth Moreover, we show, via a probabilistic formulation, how the main cause of disparity is two-fold: under/over representation and model generalization differences across subgroups.

@Aaroth It seems that the problem is deeper: Privacy attacks target subgroups disparately, regardless of whether a model is trained in a standard way, or with DP. We discuss this in "Disparate Vulnerability": arxiv.org/abs/1906.00389

Bagdasaryan and Shmatikov find that training using private SGD increases error disparities between over and under-represented groups. They blame gradient clipping, which has a larger effect on data points less like the average. An interesting fairness/privacy tradeoff.