BEGIN:VCALENDAR
VERSION:2.0
PRODID:IEEE vTools.Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:America/Los_Angeles
BEGIN:DAYLIGHT
DTSTART:20260308T030000
TZOFFSETFROM:-0800
TZOFFSETTO:-0700
RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3
TZNAME:PDT
END:DAYLIGHT
BEGIN:STANDARD
DTSTART:20261101T010000
TZOFFSETFROM:-0700
TZOFFSETTO:-0800
RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11
TZNAME:PST
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20261020T223716Z
UID:925EE77F-5E08-4BAC-94B0-B8F04CF199A7
DTSTART;TZID=America/Los_Angeles:20261020T143000
DTEND;TZID=America/Los_Angeles:20261020T153000
DESCRIPTION:Deep learning vision models are known to be vulnerable to adver
 sarial examples\, where small\, optimized input perturbations can cause co
 nfident mispredictions. A common objection is that such attacks work only 
 in the digital domain and do not survive real-world conditions. In this ta
 lk\, we show that gradient-based white-box attacks extend to deployed syst
 ems in three domains. First\, we demonstrate physical attacks on vision mo
 dels that remain effective under real-world variation in viewpoint\, light
 ing\, and distance\, including perturbations that are difficult for humans
  to perceive. Second\, we adapt the methodology to large language models\,
  producing jailbreak and prompt injection attacks that elicit harmful cont
 ent despite safety alignment. Third\, we show that an adversarial image di
 splayed on screen can hijack computer-use agents into executing attacker-c
 hosen commands. These results show that deployed AI models can be reliably
  compromised by adversaries with gradient access\, with consequences that 
 grow as models move from perception to autonomous action. We present an au
 tomated framework for systematic adversarial evaluation\, which we argue s
 hould become a standard part of AI deployment.\n\nCo-sponsored by: Santa C
 lara University \n\nSpeaker(s): \, Dr. Nageen Himayat\n\nRoom: 1302\, Bldg
 : SCDI\, 500 El Camino Real \, Santa Clara \, California\, United States\,
  95053
LOCATION:Room: 1302\, Bldg: SCDI\, 500 El Camino Real \, Santa Clara \, Cal
 ifornia\, United States\, 95053
ORGANIZER:bdezfouli@scu.edu
SEQUENCE:10
SUMMARY:Evaluating the Adversarial Robustness of Deployed AI Models
URL;VALUE=URI:https://events.vtools.ieee.org/m/587229
X-ALT-DESC:Description: <br /><p>Deep learning vision models are known to b
 e vulnerable to adversarial examples\, where small\, optimized input pertu
 rbations can cause confident mispredictions. A common objection is that su
 ch attacks work only in the digital domain and do not survive real-world c
 onditions. In this talk\, we show that gradient-based white-box attacks ex
 tend to deployed systems in three domains. First\, we demonstrate physical
  attacks on vision models that remain effective under real-world variation
  in viewpoint\, lighting\, and distance\, including perturbations that are
  difficult for humans to perceive. Second\, we adapt the methodology to la
 rge language models\, producing jailbreak and prompt injection attacks tha
 t elicit harmful content despite safety alignment. Third\, we show that an
  adversarial image displayed on screen can hijack computer-use agents into
  executing attacker-chosen commands. These results show that deployed AI m
 odels can be reliably compromised by adversaries with gradient access\, wi
 th consequences that grow as models move from perception to autonomous act
 ion. We present an automated framework for systematic adversarial evaluati
 on\, which we argue should become a standard part of AI deployment.</p>
END:VEVENT
END:VCALENDAR