My research primarily focuses on the security and privacy of machine learning. I tend to fall more on the theoretical side of things, and a lot of my recent work has focused specifically on interesting applications of differential privacy. Since 2023, I have had the honor of being advised by Kassem Fawaz as a member of the Wisconsin Privacy and Security Group.
In this work, we propose two new algorithms for privacy-preserving online learning in contexts like healthcare where people may not be comfortable sharing their personal data with a single central curator. The most natural approach here is to add independent noise to each data point before it's shared, which we show leads to two interesting downstream effects for online learning algorithms using the data. Firstly, the accumulated noise across time steps leads to a strong form of stability, which we show can be leveraged to provide amplified privacy guarantees at essentially no loss in performance. Secondly, we show that privatizing our data in this way makes it possible to simultaneously A) fit robust ML models for time-series prediction to the noisy data and B) use the experts framework to privately identify the best-performing model in real-time, all at no additional privacy cost! This approach turns out to be quite powerful, to such an extent that our local DP algorithm is able to outperform SOTA central DP baselines on a real-world prediction task using data from the COVID-19 pandemic.
(Original Paper) (Poster) (Video) (Github)
"Private Continual Counting" refers to the problem of computing private estimates for every partial sum of a sequence at once. It's attracted a lot of attention in the last few years because of its applications in machine learning (e.g. in the DP-FTRL algorithm). Our paper addresses a major limitation of prior algorithms for private counting, which is that they are all bounded: their privacy guarantees only hold when the full input size is known in advance. Building on a cool duality between lower-triangular Toeplitz matrices and generating functions (inspired by Dvijotham et al.), we design and analyze the first algorithm for private continual counting which can guarantee privacy and nearly-optimal utility even for inputs of unbounded size.
You can also read about the projects I worked on as an undergraduate research assistant at the University of Arizona (which somehow feels like a lifetime ago):
My work with Dr. Debray focused on understanding the problem of authorship attribution for programs, and reckoning with the surprising fact that it seems to be possible to figure out who wrote a program solely using stylistic features from the compiled binary. After reproducing a state-of-the-art system for binary classification, we developed a novel gray-box technique for obscuring an author's stylistic fingerprint using non-standard compiler optimizations. Our paper presenting our results and discussing the current state of binary stylometry was accepted for presentation at the 2021 CheckMATE Workshop.
(Slides) | (Talk) | (Website) | (Original Paper) | (User Study)
MetroSets is an online platform for set visualization which I helped to create as part of my research with Dr. Kobourov. The basic idea is to borrow the visual language of metro maps and apply it to abstract data, with sets drawn as metro lines and elements drawn as stations. Here's an example, visualizing the cards used in Magic: The Gathering tournament decks (new tab to zoom):
Our paper, MetroSets: Visualizing Sets as Metro Maps, was presented at IEEEVis 2020, and published in IEEE Transactions on Visualization and Computer Graphics, which is the premier journal in visualization. Subsequently, we worked on empirically evaluating our system through a controlled user study. Our paper On the Readability of Abstract Set Visualizations describing our results and situating MetroSets in the broader ecosystem of set visualization tools was presented at IEEE PacificVis and published in IEEE TVCG.
In recognition of my work on MetroSets, I was selected for honorable mention for the Computing Research Association's Outstanding Undergraduate Researcher Award for 2021, as well as the University of Arizona College of Science Outstanding Undergraduate Researcher Award. The latter award is given to only one student in the College of Science each year.