BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.4.0.1//EN TZID:Europe/Paris X-WR-TIMEZONE:Europe/Paris BEGIN:VEVENT UID:244@isdm.umontpellier.fr DTSTART;TZID=Europe/Paris:20241209T140000 DTEND;TZID=Europe/Paris:20241209T140000 DTSTAMP:20260825T134428Z URL:https://isdm.umontpellier.fr/events/avoiding-pitfalls-for-privacy-acco unting-of-dp-sgd/ SUMMARY:Avoiding Pitfalls for Privacy Accounting of DP-SGD DESCRIPTION:Room 02.124\, Building 5\, St Priest campus\n\nMachine Learning in Montpellier\, Theory &\; Practice\n\nIn this talk\, we present some technical details of privacy accounting for differentially private mechan isms. We focus on privacy accounting for Differentially Private Stochastic Gradient Descent (DP-SGD). First\, we introduce various techniques used t o analyze the privacy guarantees of DP-SGD. Then\, we discuss two differen t batch generation methods. Most privacy accounting techniques assume that batches for DP-SGD are generated using Poisson subsampling. However\, Poi sson subsampling is often not feasible in practice when training models on large datasets. We can instead implement a different technique such as sa mpling without replacement. The choice of sampling scheme has typically be en regarded as a minor technical detail. We demonstrate that the sampling scheme significantly impacts the privacy guarantees for some (realistic) D P-SGD hyperparameters.\n\nMachine Learning in Montpellier\, Theory &\; Practice ATTACH;FMTTYPE=image/jpeg:https://isdm.umontpellier.fr/wp-content/uploads/ 2026/06/ml-mtp-gC78d5.png CATEGORIES:ML MTP LOCATION:Room 02.124 Building 5\, St Priest Campus\, Montpellier\, 34000\, France X-APPLE-STRUCTURED-LOCATION;VALUE=URI;X-ADDRESS=St Priest Campus\, Montpell ier\, 34000\, France;X-APPLE-RADIUS=100;X-TITLE=Room 02.124 Building 5:geo :0,0 END:VEVENT BEGIN:VTIMEZONE TZID:Europe/Paris X-LIC-LOCATION:Europe/Paris BEGIN:STANDARD DTSTART:20241027T020000 TZOFFSETFROM:+0200 TZOFFSETTO:+0100 TZNAME:CET END:STANDARD END:VTIMEZONE END:VCALENDAR