« All Events

  • This event has passed.

AMS Special Seminar Series | Alireza Fallah

February 7 @ 12:00 pm - 1:00 pm

Location: AMS Conf. Rm N425

When: February 7th at 12:00 p.m.

Title: Machine Learning: Algorithmic and Economic Perspectives

Abstract: Data fuels modern machine learning models and algorithms, but its widespread collection has raised increasing concerns about privacy and social welfare. In response, the landscape of data acquisition and marketplaces has rapidly evolved, with tech companies implementing privacy mechanisms and governments enforcing various regulations. In this talk, I will explore the growing challenges in data privacy, spanning economic incentives, policy considerations, and algorithmic trade-offs.

The first part of the talk investigates a data marketplace involving users, platforms, and data buyers. Users benefit from platform services in exchange for data, incurring privacy loss when their data, albeit noisily, is shared with the buyer. The user chooses platforms to share data with, while platforms decide on data privacy levels and pricing before selling to the buyer. The buyer finally selects platforms to purchase data from. Using a multi-stage game-theoretic framework, I demonstrate how platform competition and buyer valuation shape user participation, platform viability, and overall welfare. I also discuss privacy regulatory interventions that can enhance user utility in mixed markets of high- and low-cost platforms.
The second part of the talk addresses another pressing challenge: heterogeneity in user privacy sensitivity across different data features. Traditional Local Differential Privacy (LDP) applies uniform protection, even to less sensitive features, often degrading the utility of downstream tasks. To overcome this, we introduce Bayesian Coordinate Differential Privacy (BCDP)—a framework that tailors privacy protection based on feature sensitivity, improving utility while preserving strong privacy guarantees. We characterize BCDP’s properties, establish its connections to standard privacy frameworks, and demonstrate its advantages in private mean estimation and ordinary least-squares regression, where it outperforms LDP in accuracy without compromising privacy.

Bio: Alireza Fallah is a postdoctoral researcher at UC Berkeley, hosted by Michael Jordan. In the summer of 2023, he obtained his Ph.D. in Electrical Engineering and Computer Science from MIT, where he worked with Asu Ozdaglar and Daron Acemoglu. He spent the fall of 2023 as the Gamelin Postdoctoral Fellow at the Simons Laufer Mathematical Sciences Institute (formerly MSRI), where he was a member of the Mathematics and Computer Science of Market and Mechanism Design program. He has received a number of awards and fellowships, including the honorable mention at the ACM SIGecom Doctoral Dissertation Award, the Ernst A. Guillemin MIT M.Sc. Thesis Award, the Apple Scholars in AI/ML Ph.D. Fellowship, the MathWorks Engineering Fellowship, and the Siebel Scholarship.

Zoom link: https://wse.zoom.us/j/92755277282?pwd=iULpLaFnWAcWl6tQYUbeyZaN3zwBzn.1

Details

Date:
February 7
Time:
12:00 pm - 1:00 pm
Event Category:

Venue

Wyman N425
3400 North Charles Street
Baltimore, 21218 United States + Google Map