db-d2
> privacy research · offensive security
R2-D2 style droid, David Benson's mark

David Benson

Cybersecurity & Data Science at Bloomberg LP / BA in Data Science, Columbia University

About

I work in cybersecurity and data science at Bloomberg, at the point where machine learning and security overlap. I recently earned my BA in Data Science from Columbia, where my research asked how much generative models memorize their training data, how to measure that honestly, and why it's so hard to remove once it's there. Off the clock, I break into things on Hack The Box.

Research

Membership inference · privacy evaluation

Fishing for Privacy: Machine Unlearning for Single-Cell VAEs

David Benson, Columbia University

VAEs trained on gene-expression data memorize rare biological subpopulations — exactly the patients a membership-inference attack can single out. Whether that memory can be scrubbed without retraining hinges first on the measurement: standard MIA protocols draw their controls from the wrong cell types and mistake subtype identity for training membership. Corrected with within-subtype matching and a trained attacker on model-internal features, the real signal is small and coherent — and under that lens ten unlearning methods all fail on structured forget sets across three datasets. The reason is structural: a VAE's shared decoder entangles a forget set's influence with the rest of the model far more than a classifier does, so it can't be selectively removed. If privacy actually matters, retrain.

0.77 → 0.53 MIA AUC after within-subtype correction 17× decoder Fisher overlap vs. classifier 10 methods, all fail
Read the full study →

Security write-ups

Hack The Box machine walkthroughs — recon to root, with the reasoning shown, not just the commands.

Hack The Box write-ups   db-d2.github.io →

Elsewhere