BEGIN:VCALENDAR VERSION:2.0 X-WR-CALNAME:EventsCalendar PRODID:-//hacksw/handcal//NONSGML v1.0//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/New_York LAST-MODIFIED:20240422T053451Z TZURL:https://www.tzurl.org/zoneinfo-outlook/America/New_York X-LIC-LOCATION:America/New_York BEGIN:DAYLIGHT TZNAME:EDT TZOFFSETFROM:-0500 TZOFFSETTO:-0400 DTSTART:19700308T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU END:DAYLIGHT BEGIN:STANDARD TZNAME:EST TZOFFSETFROM:-0400 TZOFFSETTO:-0500 DTSTART:19701101T020000 RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT CATEGORIES:College of Engineering,Thesis/Dissertations DESCRIPTION:Faculty Supervisor: Long Jiao, Computer & Information Science Committee Members: Dr. Joshua Carberry, Computer & Information ScienceDr. Lance Fiondella, Electrical & Computer Engineering Abstract: Large vision -language models rely on pretrained vision encoders to translate images in to feature representations used by downstream language models. This create s a security risk when the encoder is compromised by a stealthy backdoor a ttack, such as BadVision, where a subtle trigger causes an image to be map ped toward an attacker-chosen target representation while clean inputs rem ain largely unaffected. Because the model behaves normally under standard evaluation, these attacks are difficult to detect. This thesis investigate s controlled noise injection as a lightweight input-side defense against B adVision-style backdoors. The proposed approach adds small perturbations t o input images before they enter the vision encoder, with the goal of disr upting the trigger while preserving the semantic content of clean images. Several perturbation types are evaluated, including Gaussian noise, random noise, salt-and-pepper noise, low-frequency noise, geometric transformati ons, occlusion, scaling, rotation, and channel-based distributions. Experi mental results show that geometric and channel-based transformations have limited effect on the backdoor, while pixel-level statistical perturbation s significantly reduce target similarity, increase feature-space distance from the attacker’s target representation, and lower attack success. The se findings suggest that stealthy encoder-level triggers depend on fragile statistical patterns and can be weakened through controlled noise injecti on without requiring retraining of the full multimodal model. For furt her information please contact Dr. Long Jiao at ljiao@umassd.edu\nEvent pa ge: https://www.umassd.edu/events/cms/a-noise-based-defense-for-stealthy-b ackdoor-attacks-in-large-vision-language-models.php X-ALT-DESC;FMTTYPE=text/html:

Faculty Supervisor: Long Jiao\, Computer & Information Science

Committee Members:

Dr . Joshua Carberry\, Computer & Information Science
Dr. Lance Fiondell a\, Electrical & Computer Engineering

Abstract: Large vision- language models rely on pretrained vision encoders to translate images int o feature representations used by downstream language models. This creates a security risk when the encoder is compromised by a stealthy backdoor at tack\, such as BadVision\, where a subtle trigger causes an image to be ma pped toward an attacker-chosen target representation while clean inputs re main largely unaffected. Because the model behaves normally under standard evaluation\, these attacks are difficult to detect. This thesis investiga tes controlled noise injection as a lightweight input-side defense against BadVision-style backdoors. The proposed approach adds small perturbations to input images before they enter the vision encoder\, with the goal of d isrupting the trigger while preserving the semantic content of clean image s. Several perturbation types are evaluated\, including Gaussian noise\, r andom noise\, salt-and-pepper noise\, low-frequency noise\, geometric tran sformations\, occlusion\, scaling\, rotation\, and channel-based distribut ions. Experimental results show that geometric and channel-based transform ations have limited effect on the backdoor\, while pixel-level statistical perturbations significantly reduce target similarity\, increase feature-s pace distance from the attacker’s target representation\, and lower atta ck success. These findings suggest that stealthy encoder-level triggers de pend on fragile statistical patterns and can be weakened through controlle d noise injection without requiring retraining of the full multimodal mode l.

For further information please contact Dr. Long Jiao at ljiao@umassd.edu

Event page: https://www.umassd.edu/events/cms/a-noise-based-defens e-for-stealthy-backdoor-attacks-in-large-vision-language-models.php

DTSTAMP:20260505T154344 DTSTART;TZID=America/New_York:20260525T120000 DTEND;TZID=America/New_York:20260525T130000 LOCATION:Dion 311 SUMMARY;LANGUAGE=en-us:A Noise-Based Defense for Stealthy Backdoor Attacks in Large Vision-Language Models UID:c45be63911af2bf51a7dfe92826e274a@www.umassd.edu END:VEVENT END:VCALENDAR