While several results in the literature demonstrate that Bayesian inference approximated by MCMC output can achieve differential privacy with zero or limited impact on the ensuing posterior, we reassess this perspective via an alternate “exact" MCMC perturbation within a federated learning setting. Since the ensuing privacy criterion is mostly related to a slowing-down of MCMC convergence rather than a generic gain in protecting data privacy, we propose an alternative decision-theoretic framework that accommodates more realistic privacy constraints.
19 Gennaio 2026, ore 10.30
Christian Robert
Université Paris Dauphine, PSL & University of Warwick
In person: Room V (4th floor) building CU002 Scienze Statistiche
Webinar: https://uniroma1.zoom.us/j/83625004899?pwd=bXCtz0mp759PUh2lkqT0BUoVa0Uegg.1
ID riunione: 836 2500 4899
Passcode: 123456