Join Klaus Mueller, professor of computer science and interim chair of the Department of Technology and Society, as he hosts Sucheta Lahiri.

Lahiri leads the AI Ethics and Risk Management function at Oxy, where she is responsible for ensuring that the company's AI solutions are developed and deployed in a manner that is ethical, efficient, trustworthy, safe, sustainable, and human-centered. She holds a doctorate from Syracuse University, along with two master's degrees in Applied Statistics and Information Science earned in India.

Zoom: https://stonybrook.zoom.us/j/7851507944?omn=98268154363#success

Abstract: Recent studies have highlighted the vulnerability of Natural Language Processing (NLP) and Vision-Language Models (VLMs) to backdoor attacks, posing significant security risks. Understanding these attack strategies is crucial for assessing model robustness and developing effective defenses. This thesis proposal aims to investigate the vulnerability of language and vision-language models, analyze abnormal behaviors in backdoor-attacked models, and develop defense methods to enhance safety of modern machine learning models at deployment.


We investigate the internal mechanisms of backdoored NLP models, identifying a distinct attention focus drifting phenomenon, where trigger tokens hijack attention regardless of the input context. Through comprehensive qualitative and quantitative analysis, we provide insights into the underlying mechanisms that enable backdoor attacks. Building on these insights, we propose detection methods to differentiate backdoored models from clean ones, through inspecting both the attention distribution and the model predictions. To better understand the vulnerability, we develop advanced backdoor attack strategies targeting language models in classification tasks. For BERT variants, we introduce Trojan Attention Loss (TAL), a novel method that directly manipulates attention patterns to enhance backdoor effectiveness, ensuring stealth and robustness. Vision-Language Models have demonstrated strong performance in recent years. Yet their vulnerability is largely underexplored. We investigate advanced backdoor attack strategies on Vision-Language Models, focusing on image-to-text generation tasks. We demonstrate how backdoors can be embedded in complex multimodal tasks while maintaining semantic integrity under poisoned inputs. Additionally, we propose innovative techniques for injecting backdoors without requiring access to the original training data, expanding the feasibility of real-world attacks.

This proposal provides novel insights into the internal mechanisms of backdoored models, propose effective detection strategies, and develop advanced attack techniques that expose critical vulnerabilities. These findings underscore the urgent need for robust security measures to defend against emerging backdoor threats in deep learning models. The results have been published in top venues including ICLR, ECCV, NAACL, EMNLP, etc.

Speaker: Weimin Lyu


Zoom link: https://stonybrook.zoom.us/j/99880605139?pwd=cfWbRG6n9v3GXEa7OqvXa5cOp5eLBv.1
Meeting ID: 998 8060 5139
Passcode: 843302
Speaker Petar Djuric Refreshments will be provided Deep Gaussian processes: Theory and applications Petar M. Djurić Department of Electrical and Computer Engineering Stony Brook University Abstract: Gaussian processes are an infinite-dimensional generalization of multivariate normal distributions. They provide a principled approach to learning with kernel machines and they have found wide applications in many fields. More recently, with the advance of deep learning, the concept of deep Gaussian processes has emerged. Deep Gaussian processes can be viewed as multilayer hierarchical organizations of Gaussian processes that are equivalent to infinitely wide multiple layer neural networks. Deep Gaussian processes have improved capacity for prediction and classification over standard Gaussian processes, while models based on them continue to allow for full Bayesian treatment and for applications when the amount of available data is limited. The theory of recent progress in deep Gaussian processes will be presented and some applications will be provided. Biosketch: Petar M. Djurić received the B.S. and M.S. degrees in electrical engineering from the University of Belgrade, Belgrade, Yugoslavia, respectively, and the Ph.D. degree in electrical engineering from the University of Rhode Island, Kingston, RI, USA. He is a SUNY Distinguished Professor and currently, he is a Chair of the Department of Electrical and Computer Engineering, Stony Brook University, Stony Brook, NY, USA. Djurić was a recipient of the IEEE Signal Processing Magazine Best Paper Award in 2007 and the EURASIP Technical Achievement Award in 2012. From 2008 to 2009, he was a Distinguished Lecturer of the IEEE Signal Processing Society. He was the Editor-in-Chief of the IEEE Transactions on Signal and Information Processing over Networks (2015-2018). Djurić is a Fellow of IEEE and EURASIP
Join Zoom Meeting
https://stonybrook.zoom.us/j/98079526509?pwd=Wkt5eURhVDN5VE56TUloS2h5V1Jodz09

Meeting ID: 980 7952 6509
Passcode: 949941



Abstract Over the last decade, artificial neural networks have undergone a revolution, catalyzed by better
supervised learning algorithms. However, in stark contrast to young animals (including humans), training such
networks requires enormous numbers of labeled examples, leading to the belief that animals must rely instead
mainly on unsupervised learning. The reason is that most animal behavior is not the result of clever learning
algorithms--supervised or unsupervised-- but is encoded in the genome. Specifically, animals are born with
highly structured brain connectivity, which enables them to learn very rapidly. Because the wiring diagram is
far too complex to be specified explicitly in the genome, it must be compressed through a genomic
bottleneck. I will describe results showing how the genomic bottleneck algorithm can lead to dramatic
compression of networks and better generalization, particularly for rapid transfer in supervised and
reinforcement learning.

Anthony Zador is professor of neuroscience at CSHL.
Presented by Stony Brook University Department of Biomedical Informatics and Long Island Network for Clinical and Translational Science (LINCATS).

The seminar aims to empower participants with the knowledge and skills necessary to harness AI effectively in clinical practice and research. It will equip attendees with practical insights, case studies, and interactive discussions led by experts in both AI and medicine, fostering a collaborative environment where attendee can explore how to overcome barriers and maximize the potential of AI in transforming modern healthcare delivery.

All Stony Brook Audiences Welcome.
Please note: This exciting event is open to all Stony Brook Faculty/Staff/Students. While the overarching theme for this event is the application of AI in medicine, the event is designed to bridge the professional practice gap that exists between cutting-edge AI research and its practical implementation in clinical settings, While AI holds immense promise for transforming healthcare delivery, many physicians and researchers lack the foundational knowledge and practical skills needed to effectively integrate AI into their daily practices.

THIS CONFERENCE IS FOR STONY BROOK UNIVERSITY & HOSPITAL FACULTY/STAFF & STUDENTS ONLY.


Registration link: https://cme.stonybrookmedicine.edu/continuing-medical-education/conferences/235/bench-to-bedside-understanding-the-practical-application-of-ai-in-medicine-2024/10/17/2024

FOR QUESTIONS
joseph.cesaria@stonybrookmedicine.edu
mary.saltz@stonybookmedicine.edu

Time: Jan 26, 2021 03:00 PM Eastern Time (US and Canada)

All are welcome!

Zoom Meeting:
https://stonybrook.zoom.us/j/93818552212?pwd=ajZkT2x4a2tiaDJUL1h3VFhLZEgwQT09

Meeting ID: 938 1855 2212
Passcode: 802722

Title: Data-Driven Document Unwarping

Abstract: Capturing document images is a common way to digitize and record physical documents due to the ubiquitousness of mobile cameras. To make text recognition easier, it is often desirable to digitally flatten a document image when the physical document sheet is folded or curved. However, unwarping a document from a single image in natural scenes is very challenging due to the complexity of document sheet deformation, document texture, and environmental conditions. Previous model-driven approaches struggle with inefficiency and limited generalizability. In this thesis, I investigate several data-driven approaches to tackle the document unwarping problem.

Data acquisition is the central challenge in data-driven methods. I first design an efficient data synthesis pipeline based on 2D image warping and train DocUNet, the pioneering data-driven document unwarping model, on the synthetic data. A benchmark dataset is also created to facilitate comprehensive evaluation and comparison. To improve the unwarping performance by training on more realistic data, I introduce the Doc3D dataset and DewarpNet. Supervised by 3D shape ground truth in Doc3D, DewarpNet is significantly better than DocUNet. DocUNet and DewarpNet depend on the synthetic data for the ground truth deformation annotation. To exploit the real-world images, I propose PaperEdge, a weakly supervised model trained with in-the-wild document images with easy-to-obtain boundary information. PaperEdge surpasses DewarpNet by utilizing both the synthetic data and weakly annotated real data in the Document In the Wild (DIW) dataset. Finally, I propose to incorporate the 3D physical constraints in training DewarpNet and PaperEdge. The constraints regulate the possible deformations on document papers. I also propose to augment the Doc3D and DIW dataset by introducing an online document segmentation model and better hardware.