BEGIN:VCALENDAR
VERSION:2.0
PRODID:icalendar-ruby
CALSCALE:GREGORIAN
METHOD:PUBLISH
BEGIN:VTIMEZONE
TZID:Europe/Vienna
BEGIN:DAYLIGHT
DTSTART:20240331T030000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
TZNAME:CEST
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20231029T020000
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
TZNAME:CET
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260424T143107Z
UID:65e99f0286610983097176@ist.ac.at
DTSTART:20240325T130000
DTEND:20240325T140000
DESCRIPTION:Speaker: Volkan Cevher\nhosted by Marco Mondelli\nAbstract: One
  prominent approach toward resolving the adversarial vulnerability of deep
  neural networks is the two-player zero-sum paradigm of adversarial traini
 ng\, in which predictors are trained against adversarially-chosen perturba
 tions of data. Despite the promise of this approach\, algorithms based on 
 this paradigm have not engendered sufficient levels of robustness\, and su
 ffer from pathological behavior like robust overfitting. To understand thi
 s shortcoming\, we first show that the commonly used surrogate-based relax
 ation used in adversarial training algorithms voids all guarantees on the 
 robustness of trained classifiers. The identification of this pitfall info
 rms a novel non-zero-sum bilevel formulation of adversarial training\, whe
 rein each player optimizes a different objective function. Our formulation
  naturally yields a simple algorithmic framework that matches and in some 
 cases outperforms state-of-the-art attacks\, attains comparable levels of 
 robustness to standard adversarial training algorithms\, and does not suff
 er from robust overfitting.
LOCATION:Office Bldg West / Ground floor / Heinzel Seminar Room (I21.EG.101
 )\, ISTA
ORGANIZER:jdeanton@ist.ac.at
SUMMARY:Volkan Cevher: Adversarial Training Should Be Cast as a Non-Zero-Su
 m Game
URL:https://talks-calendar.ista.ac.at/events/4870
END:VEVENT
END:VCALENDAR