Calister Nnona

Calister Nnona

Machine-learning privacy and security researcher
contact [at] calisternnona [dot] com · GitHub

I work on the privacy and security of machine-learning systems, with a focus on differential privacy, synthetic data, and the gap between what a guarantee promises and what a deployment delivers.

My current project is CoRTeC, a mechanism for differentially private synthetic tabular data that spends its privacy budget once, on a statistics release, and lets a frozen language model decode it. The work measures the deployed alternatives on the questions an institution actually asks: if I train on this instead of the real data, what do I lose, and what can an adversary learn from the output?

[Add a sentence or two about your background: role, affiliation, education, and what you did before this. Edit public/index.html.]

Research

2026
CoRTeC: Cohort-Conditioned Differentially Private Synthetic Tabular Data from a Frozen Language Model
Differentially private synthetic tabular data lets an institution train models on, and share, records it cannot release. The mechanisms in use today are selected on marginal fidelity and do not report downstream utility. CoRTeC spends its budget once, on a release of cohort-conditioned histograms and a conditional table, and lets a frozen, un-finetuned language model decode it, so unlimited records cost no further budget. At ε = 2 on three regulated-domain datasets its marginal error is within 0.016 of the most accurate marginal method and tree models trained on it reach the real-sample floor. The paper also shows that both standard acceptance criteria for synthetic data are saturated.
Paper · Technical report · Research code · Reference implementations
Last updated September 2026.