BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Sabre//Sabre VObject 4.6.1//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:Europe/Zurich
X-LIC-LOCATION:Europe/Zurich
TZURL:http://tzurl.org/zoneinfo/Europe/Zurich
BEGIN:DAYLIGHT
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
DTSTART:19810329T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
TZNAME:CET
DTSTART:19961027T030000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
UID:news2093@dmi.unibas.ch
DTSTAMP;TZID=Europe/Zurich:20261102T191640
DTSTART;TZID=Europe/Zurich:20261111T161500
SUMMARY:Building neighbourhoods that keep secrets and catch trojans: privac
 y and robustness native to decentralized ML
DESCRIPTION:Abstract:Decentralized learning is an emerging ML paradigm wher
 e nodes collaboratively train models without a central server. However\, t
 his invites invites two kinds of bad neighbours. The nosy one follows the 
 protocol but mines the models it receives for private data through gradien
 t inversion or membership inference. The saboteur plants backdoors that be
 nign carriers then spread to honest nodes far from the attacker. In this t
 alk\, I argue that the neighbourhood itself can be turned into the defence
 . Against the nosy neighbour\, SHATTER (PoPETS 2025) has each node split i
 ts model into randomized chunks and scatter them through virtual nodes\, s
 o that no neighbour ever holds a full model to attack. Against the saboteu
 r\, ARGUS (NeurIPS 2026) builds on the observation that triggers extracted
  locally from benign models look random\, while those from backdoored mode
 ls follow a consistent pattern. By pooling this trigger information\, a ne
 ighbourhood can collectively expose the malicious node. An open question t
 hat naturally emergers from this: Accountability wants transparency\, and 
 privacy wants the opposite. Can DL have both by using the core ideas behin
 d SHATTER (model chunking) or ARGUS (validate using collective neighbourho
 od intelligence)?
X-ALT-DESC:<p>Abstract:<br /><br />Decentralized learning is an emerging ML
  paradigm where nodes collaboratively train models without a central serve
 r. However\, this invites invites two kinds of bad neighbours. The nosy on
 e follows the protocol but mines the models it receives for private data t
 hrough gradient inversion or membership inference. The saboteur plants bac
 kdoors that benign carriers then spread to honest nodes far from the attac
 ker. In this talk\, I argue that the neighbourhood itself can be turned in
 to the defence. Against the nosy neighbour\, SHATTER (PoPETS 2025) has eac
 h node split its model into randomized chunks and scatter them through vir
 tual nodes\, so that no neighbour ever holds a full model to attack. Again
 st the saboteur\, ARGUS (NeurIPS 2026) builds on the observation that trig
 gers extracted locally from benign models look random\, while those from b
 ackdoored models follow a consistent pattern. By pooling this trigger info
 rmation\, a neighbourhood can collectively expose the malicious node. An o
 pen question that naturally emergers from this: Accountability wants trans
 parency\, and privacy wants the opposite. Can DL have both by using the co
 re ideas behind SHATTER (model chunking) or ARGUS (validate using collecti
 ve neighbourhood intelligence)?</p>
END:VEVENT
END:VCALENDAR