Aurélien Bellet

Forget expensive multi-run privacy audits: this canary crafting method lets you estimate privacy leakage more accurately and cheaply in a single training run.

Lumberjack's heavy hitter detection algorithm lets you prune differentially private random forests with less accuracy loss, finally making them practical for sensitive tabular data.

Despite theoretical guarantees, widely-used differential privacy libraries harbor subtle bugs that can be exposed by a new gray-box auditing technique.