- Diversity, Equity, and Inclusion
- Research News
- Department Life
- Oral History of Cornell CS
- CS 40th Anniversary Booklet
- ABC Book for Computer Science at Cornell by David Gries
- Department Timeline
- Job Postings
- Ithaca Info
- Internal info
- Graduation Information
- Cornell Tech Colloquium
- Student Colloquium
- Fall 2023 Colloquium
- Conway-Walker Lecture Series
- Salton 2023 Lecture Series
- Seminars / Lectures
- Big Red Hacks
- Cornell University High School Programming Contests 2023
- Game Design Initiative
- CSMore: The Rising Sophomore Summer Program in Computer Science
- Explore CS Research
- ACSU Research Night
- Cornell Junior Theorists' Workshop
Speaker 1: Eugene Bagdasaryan
Title: Spinning Language Models for Propaganda-as-a-Service
Abstract: We investigate a new threat to generative language models: training-time attacks that cause models to “spin” their outputs so as to support an adversary-chosen sentiment or point of view — but only when the input contains adversary-chosen trigger words. For example, a spinned summarization model outputs positive summaries of articles that mention the name of some individual or organization.
Model spinning introduces a “meta-backdoor” into a model: outputs of spinned models preserve context and maintain standard accuracy metrics, yet also satisfy a meta-task chosen by the adversary. This approach enables propaganda-as-a-service — the attacker can introduce any bias for any topic into the model on demand.
We evaluate this attack on language generation, summarization, and translation models with different triggers and meta-tasks such as sentiment, toxicity, and entailment. Spinned models largely maintain their accuracy metrics while shifting their outputs to satisfy the adversary’s meta-task. Finally, we propose a black-box, meta-task-independent defense that, given a list of candidate triggers, can detect models that selectively apply spin to inputs.
Bio: Eugene is a PhD candidate in Computer Science at Cornell Tech, Cornell University. He studies machine learning privacy and security under real-world settings. He works on understanding how ML-based systems can fail or cause harm and how to make these systems better.
Speaker 2: Utkarsh Mall
Title: Visual Discovery from Spatio-temporal Imagery
Abstract: From social media all the way to satellite images, we are capturing visual data at an unprecedented scale. These images tell a story about our planet. With advances in automatic recognition, we can build a collective understanding of world-scale events as recorded through visual media. In GeoStyle, we use supervised vision on spatio-temporal social media images to understand fashion trends and discover cultural phenomena and social events around the world. Broadening our domain to include satellite imagery, we introduce completely unsupervised techniques to discover interesting change events across the planet from satellite images. Our framework can be potentially applied in different visual domains ranging from sustainability to online commerce to discover interesting phenomena in those domains.
Bio: Utkarsh Mall is a PhD candidate in Computer Science at Cornell University advised by Kavita Bala and Bharath Hariharan. His research interest lies in computer vision. His research focuses on building recognition models that can learn with little to no supervision and using these models to make discoveries from visual data. He has applied this work to a range of application domains from fashion to satellite images. He obtained his bachelor’s degree from IIT Bombay.
Speaker 3: Guandao Yang
Title: Deep Learning for Geometric Data Processing
Abstract: Geometry processing involves acquisition and manipulation of geometric data. One example of a geometric processing task is producing 3D model from scanned point clouds. Since such point clouds are sparse and don't represent the full surface, we are solving an under-constrained problem. As a result, we need to leverage some prior knowledge to perform well.
In this talk, I will talk about my research using deep learning to encode such prior knowledge. I will introduce my works using advance in deep generative modeling to represent and create 3D shapes (Yang et al., 2019, Cai et al. 2020). These 3D generative models allow us to acquire and complete geometric data from observations in high quality.
Bio: Guandao Yang is a Ph.D. at Cornell Tech, advised by Prof. Serge Belongie and Prof. Bharath Hariharan. His research focuses on automating geometry processing tasks using deep learning as prior and as representations. During his Ph.D., he interned at Google, Intel, and NVIDIA. Before Cornell Tech, he spent his undergrad studying Math and CS at Cornell University in Ithaca. In his spare time, he likes traditional rock climbing and piano.