BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Institute of Mathematics and Informatics - ECPv6.0.8//NONSGML v1.0//EN
CALSCALE:GREGORIAN
METHOD:PUBLISH
X-ORIGINAL-URL:https://math.bas.bg
X-WR-CALDESC:Събития за Institute of Mathematics and Informatics
REFRESH-INTERVAL;VALUE=DURATION:PT1H
X-Robots-Tag:noindex
X-PUBLISHED-TTL:PT1H
BEGIN:VTIMEZONE
TZID:Europe/Sofia
BEGIN:DAYLIGHT
TZOFFSETFROM:+0200
TZOFFSETTO:+0300
TZNAME:EEST
DTSTART:20230326T010000
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0300
TZOFFSETTO:+0200
TZNAME:EET
DTSTART:20231029T010000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTART;TZID=Europe/Sofia:20230822T160000
DTEND;TZID=Europe/Sofia:20230822T173000
DTSTAMP:20260406T150156
CREATED:20230821T171909Z
LAST-MODIFIED:20230821T171909Z
UID:14881-1692720000-1692725400@math.bas.bg
SUMMARY:Семинар на секция "Математически основи на информатиката"
DESCRIPTION:На 22 август 2023 г. (вторник) от 16:00 часа в зала 503 на ИМИ-БАН\nще се проведе заседание на семинара на секция\n„Математически основи на информатиката”. \nДоклад на тема: \nData Attribution: Scaling up and Applications to Defending Against Backdoor Attacks\nще изнесе\nКристиян Георгиев\, докторант в MIT под ръководството на Александър Мондри.\nАбстракт. The goal of data attribution is to trace model predictions back to training data. Despite a long line of work towards this goal\, existing approaches to data attribution tend to force users to choose between computational tractability and efficacy. That is\, computationally tractable methods can struggle with accurately attributing model predictions in non-convex settings (e.g.\, in the context of deep neural networks)\, while methods that are effective in such regimes require training thousands of models\, which makes them impractical for large models or datasets. \nIn this talk\, I will first introduce TRAK (Tracing with the Randomly-projected After Kernel)\, a data attribution method that is both effective and computationally tractable for large-scale\, differentiable models. Next\, I will showcase how to leverage data attribution methods for building defenses against backdoor attacks. In a backdoor attack\, an adversary inserts maliciously constructed backdoor examples into a training set to make the resulting model vulnerable to manipulation. I will show that without structural information about the training data distribution\, backdoor attacks are indistinguishable from naturally-occurring features in the data–and thus impossible to “detect” in a general sense. With this insight in mind\, I will describe how we use TRAK (and other data attribution methods) towards finding features corresponding to the “backdoored” samples.
URL:https://math.bas.bg/event/%d1%81%d0%b5%d0%bc%d0%b8%d0%bd%d0%b0%d1%80-%d0%bd%d0%b0-%d1%81%d0%b5%d0%ba%d1%86%d0%b8%d1%8f-%d0%bc%d0%b0%d1%82%d0%b5%d0%bc%d0%b0%d1%82%d0%b8%d1%87%d0%b5%d1%81%d0%ba%d0%b8-%d0%be%d1%81%d0%bd%d0%be-18/
LOCATION:Институт по математика и информатика – БАН\, Block 8\, 1113 БАН IV км.\, София\, Bulgaria
CATEGORIES:Редовен семинар
END:VEVENT
END:VCALENDAR