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:Advisor: Dr. Long Jiao, Computer & Information ScienceÌýÌýCommi ttee Members: Dr Joshua Carberry, Computer & Information Science Dr Amir Akhavan Masoumi, Computer & Information Science Abstract: AI agents have seen rapid adoption in recent years, and many of them now include a memory database that allows them to store information about their users and refe rence it across separate conversations. However, this feature also creates a new attack surface, where adversaries can inject poisoned memories into the database of an agent in order to degrade its performance. Previous re search in this area has focused on attacks that either insert poisoned mem ories directly into the memory store or assume that the attacker knows in advance which questions the agent will be asked. This thesis attempts to p oison a memory database without either of those capabilities, relying only on conversation history that has already been stored in memory and on kno wledge of the type and retrieval policy of the memory system. From this co rpus, it selects up to k conversations using closeness to the hubs of the memory database — regions of the embedding space where many unrelated us er queries tend to retrieve from — as the guiding selection criterion. C onversations chosen in this way are disproportionately retrieved across a wide range of future queries, displacing legitimate context and broadly de grading the agent's responses. Results show that this is an effective way to reduce overall AI agent performance with minimal information about the agent itself. ÌýThis research demonstrates that the embedding geometry of retrieval-based memory systems is itself a vulnerability: an adversary can exploit the structure of the embedding space without ever needing to know what the agent will be asked.ÌýÌýFor further information please contact D r. Long Jiao at ljiao@umassd.edu\nEvent page: /event s/cms/hub-based-memory-poisoning-query-blind-attacks-on-retrieval-augmente d-llm-agents.php X-ALT-DESC;FMTTYPE=text/html:

91É«ÇéƬ

Advisor: Dr. Long Jiao\, Comput er & Information ScienceÌý
Ìý
Committee Members:

\n\n

Abstract:

\n

AI agents have seen rapid adoption in recent years\, and many of them now include a memory database that allows them to store information about their users and reference it across separate conversations. However\, thi s feature also creates a new attack surface\, where adversaries can inject poisoned memories into the database of an agent in order to degrade its p erformance. Previous research in this area has focused on attacks that eit her insert poisoned memories directly into the memory store or assume that the attacker knows in advance which questions the agent will be asked. Th is thesis attempts to poison a memory database without either of those cap abilities\, relying only on conversation history that has already been sto red in memory and on knowledge of the type and retrieval policy of the mem ory system. From this corpus\, it selects up to k conversations using clos eness to the hubs of the memory database — regions of the embedding spac e where many unrelated user queries tend to retrieve from — as the guidi ng selection criterion. Conversations chosen in this way are disproportion ately retrieved across a wide range of future queries\, displacing legitim ate context and broadly degrading the agent's responses. Results show that this is an effective way to reduce overall AI agent performance with mini mal information about the agent itself. ÌýThis research demonstrates that the embedding geometry of retrieval-based memory systems is itself a vulne rability: an adversary can exploit the structure of the embedding space wi thout ever needing to know what the agent will be asked.Ìý
Ìý
Fo r further information please contact Dr. Long Jiao at ljiao@umassd.edu

Event page: http s://www.umassd.edu/events/cms/hub-based-memory-poisoning-query-blind-attac ks-on-retrieval-augmented-llm-agents.php

DTSTAMP:20260519T072243 DTSTART;TZID=America/New_York:20260520T120000 DTEND;TZID=America/New_York:20260520T130000 LOCATION:Dion 311 SUMMARY;LANGUAGE=en-us:Hub-Based Memory Poisoning: Query-Blind Attacks on R etrieval-Augmented LLM Agents UID:d293d0ba1ee8c471203de19a694dadd9@www.umassd.edu END:VEVENT END:VCALENDAR