BEGIN:VCALENDAR
VERSION:2.0
PRODID:icalendar-ruby
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VTIMEZONE
TZID:Europe/Vienna
BEGIN:DAYLIGHT
DTSTART:20260329T030000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20251026T020000
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260424T181700Z
UID:68f0cc9cb8ff1474414207@ist.ac.at
DTSTART:20251028T104500
DTEND:20251028T120000
DESCRIPTION:Speaker: Edwige Cyffers\nAbstract: Decentralized Learning (DL) 
 enables users to collaboratively train models without sharing raw data by 
 iteratively averaging local updates with neighbors in a network graph. Thi
 s setting is increasingly popular for its scalability and its ability to k
 eep data local under user control. Strong privacy guarantees in DL are typ
 ically achieved through Differential Privacy (DP)\, with results showing t
 hat DL can even amplify privacy by disseminating noise across peer-to-peer
  communications.Yet in practice\, the observed privacy-utility trade-off o
 ften appears worse than in centralized training\, which may be due to limi
 tations in current DP accounting methods for DL. In this paper\, we show t
 hat recent advances in centralized DP accounting based on Matrix Factoriza
 tion (MF) for analyzing temporal noise correlations can also be leveraged 
 in DL. By generalizing existing MF results\, we show how to cast both stan
 dard DL algorithms and common trust models into a unified formulation. Thi
 s yields tighter privacy accounting for existing DP-DL algorithms and prov
 ides a principled way to develop new ones. To demonstrate the approach\, w
 e introduce MAFALDA-SGD\, a gossip-based DL algorithm with user-level corr
 elated noise that outperforms existing methods on synthetic and real-world
  graphs.
LOCATION:Central Bldg / O1 / Mondi 2a (I01.O1.008)\, ISTA
ORGANIZER:achaturv@ist.ac.at
SUMMARY:Edwige Cyffers: TCS Seminar - Unified Privacy Guarantees for Decent
 ralized Learning via Matrix Factorization
URL:https://talks-calendar.ista.ac.at/events/6081
END:VEVENT
END:VCALENDAR