Writing

Long-form technical pieces on ML privacy, membership inference, retrieval systems, and the engineering problems behind privacy-sensitive AI. Four to eight pieces per year. RSS feed.

  1. Membership Inference at the Threshold: Likelihood Ratios and What They Expose

    Training samples sit closer to the decision boundary than non-members — here is what that means in practice, and what the ROC curve actually tells you about your model's privacy surface.

  2. Multi-Task Learning as Inductive Bias Under Distribution Shift

    Four controlled ablations that isolate when multi-task learning helps, when it hurts, and why — including the data-duplication control that separates task interaction from raw data volume.