Ai

Download

ai@lists.engr.oregonstate.edu

March 2025

2 participants
2 discussions

AI Seminar: Daniel Lowd
by Tadepalli, Prasad 13 Mar '25

13 Mar '25

Hi all, Our last AI seminar for this term is on Friday, March 14th. Title: The Manchurian Classifier: Invisible Triggers and Vulnerabilities in Text Models<https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fengineeri…> Speaker: Daniel Lowd, Professor of Computer Science, University of Oregon Time: 2:00 PM Location: KEC 1001 and Zoom Zoom link: https://oregonstate.zoom.us/s/98357211915<https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Foregonsta…> Abstract: Machine learning models are increasingly deployed in critical applications, yet their vulnerability to data poisoning remains a significant security concern. In this talk, I will explore backdoor attacks on text classifiers - a form of adversarial manipulation where hidden "trigger patterns" are embedded during training, causing models to behave normally on standard inputs but produce targeted misclassifications when triggers are present. I will first introduce the concept of backdoor attacks in the text domain and demonstrate why they represent a unique challenge compared to traditional adversarial examples. Then, I will present LLMBkd, our method that leverages large language models to insert backdoor triggers as subtle stylistic variations in text. These attacks are particularly concerning as they require minimal technical expertise while offering attackers a wide range of possible trigger mechanisms that can bypass conventional inspection. Building on this foundation, I will discuss AttrBkd, our more advanced approach that achieves even greater subtlety by extracting fine-grained attributes from existing backdoor attacks. Our comprehensive human evaluations reveal a critical insight: these attacks remain highly effective while being virtually indistinguishable from normal text to human reviewers. This disconnect between human perception and automated detection metrics exposes fundamental limitations in current defensive measures. Finally, I'll discuss the current state of defenses against backdoor attacks and broader security implications for machine learning systems that process text. If time permits, I'll also introduce our ongoing work on Malicious Programming Prompt (MaPP) attacks, which demonstrate how even state-of-the-art coding assistants can be manipulated to produce security vulnerabilities through carefully crafted prompts Biography: Daniel Lowd is a Professor of Computer Science at the University of Oregon. His research interests include machine learning and artificial intelligence, focusing on adversarial methods and statistical relational models. He received his Ph.D. in 2010 from the University of Washington. He has received a Google Faculty Award, an ARO Young Investigator Award, and a best paper award from DEXA. AI Seminars Link: https://engineering.oregonstate.edu/EECS/research/AI-seminars

1 0

AI Seminar: Human-Robot Collaboration for Everyday Household Tasks
by Abou Mrad, Christian 06 Mar '25

06 Mar '25

Dear all, Our next AI seminar is scheduled for Friday, March 7th. Talk details: AI Seminar: Human-Robot Collaboration for Everyday Household Tasks Speaker: Jacob Krantz, Research Scientist, Meta Time: 2:00 PM Location: KEC 1001 and Zoom Zoom link: https://oregonstate.zoom.us/s/98357211915<https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Foregonsta…> Talk Abstract: I want a robot assistant that can work with me to accomplish everyday tasks around the home: tidying up the living room, preparing the table for dinner, etc. This broad capability involves the use of natural language for task specification, embodied multi-agent planning, and robust skill execution. Toward this end, I will present our simulation benchmark called PARTNR: Planning And Reasoning Tasks in humaN-Robot collaboration. PARTNR stands as the largest benchmark of its kind, comprising 100,000 natural language tasks spanning 60 houses and 5,819 unique objects. We analyze state-of-the-art LLMs as planners for these tasks, and reveal limitations in coordination, task tracking, and failure recovery. When paired with real human partners, these LLMs are less efficient than human-human collaboration by 1.5x and less efficient than a single human by 1.1x. Through the PARTNR benchmark, many directions of research can be pursued, for example multi-agent planning, 3D scene understanding, skill coordination, Sim2Real transfer, and even HRI studies. Speaker Bio: Jacob Krantz is a research scientist at the Fundamental AI Research lab at Meta (FAIR), where he works on systems of embodied intelligence with applications toward socially intelligent robotics. In the past he co-organized several community challenges relating to semantically driven navigation (VLN, ImageNav). He received his PhD in Computer Science from Oregon State University in 2023 under Dr. Stefan Lee and his BS in Computer Science with a minor in Physics from Gonzaga University in 2019. For future AI seminars, please visit: https://engineering.oregonstate.edu/EECS/research/AI-seminars<https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fengineeri…>. Best, Christian Abou Mrad Graduate Student Oregon State University

1 1