Casey Meehan Email & Phone Number

San Francisco, California, United States

San Francisco Bay Area

Greater San Diego Area

San Francisco, California, United States

I joined with the broad goal of assessing whether self-supervised vision models (e.g. VICReg, SimCLR, Dino) memorize their training images. This has serious privacy implications since it opens the risk that images could be extracted from a trained model. However, how to define memorization for SSL models, how to quantify it, and how to demonstrate that such memorization could lead to leaking images were all open questions. Over the next several months, I experimented with training SSL models under a variety of different settings and identified a test that could detect such memorization. Then, building on research from my collaborators, I implemented a diffusion-based data reconstruction attack that could reverse-engineer training images given the model weights. Finally, we studied how such memorization responds to different training parameters to propose memorization mitigation strategies. To learn more about this work, please refer to our paper: https://arxiv.org/abs/2304.13850

Tumult Labs is making differential privacy (DP) accessible with their own platform offering DP data analysis and privacy accounting (quantification of privacy risk). Their clients in many cases have to release huge numbers of queries from massive datasets, which can lead to a high risk of exposing data if not careful. To address this problem, I worked with Tumult scientists to design new privacy accounting techniques and privacy-preserving algorithms based on findings in recent DP literature. Over several weeks of iteration I worked to design these provably private algorithms, implemented methodical tests on representative datasets, and presented baseline comparisons. Ultimately, my work demonstrated a potential technical direction Tumult leadership could consider to continue offering the premier privacy-preserving analysis platform.

Palo Alto, Ca

The Tesla sensing team wanted to design an in-house ultrasonic sensing system for running lab tests and for potential integration into vehicles. I joined to build this system from the ground up. That meant interfacing with chip vendors to evaluate the potential of their ultrasonic sensing products, electronic testbench construction, and algorithm design to reliably compute distances from the Tesla body using ultrasonics. Finally, we made risk assessments of the challenges of using an in-house platform. I left behind a testbench, a codebase, and documentation for the sensing engineers to use and build on going forward.

Norwood, Ma

I was a part of an analog design team that worked in the automotive division and then the audio division. The team focused on sigma-delta ADC design --- a broadly useful architecture for many application spaces. I designed, simulated, verified, and oversaw the layout of several blocks in an (at the time) SOTA 40nm process. Our designs are now in production as part of a top-tier acoustic noise cancellation ASIC. I learned a tremendous amount from my highly skilled management there: everything from signal processing theory to simulation technique from lab test methods to parking lot basketball plays.

B.S., Electrical Engineering & Signal Processing

M.S., Computational Science And Engineering

Doctor Of Philosophy - Phd, Computer Science

My work focused on privacy in various ML settings, both from an empirical perspective (model memorization and inversion attacks).