Differential privacy for aggregates - add dp_laplace and dp_gaussian built-in functions with budget accounting#2539

Open

andersonm-ibm wants to merge 26 commits into

apache:mainfrom

andersonm-ibm:diff_privacy_pub

Commits on Jul 7, 2026

Intro notebooks
Maya Anderson
committed
Add CLAUDE.md
Maya Anderson
committed
Initial DPBuiltinCPInstruction and RDPAccountant
Maya Anderson
committed
Add dp_gaussian and dp_laplace operations
Maya Anderson
committed
CLAUDE_CODE_PLAN.md
Maya Anderson
committed
Fix test package name.
Maya Anderson
committed
Fix Laplace accountant
Maya Anderson
committed
Add unit tests
Maya Anderson
committed
Rename RDPAccountant to DPBudgetAccountant
Maya Anderson
committed
Delete RDPAccountant
Maya Anderson
committed
Fix parseInstruction() comment
Maya Anderson
committed
Integration tests
Maya Anderson
committed
Update test comments
Maya Anderson
committed
Fix integration test to compare clean to noisy data
Maya Anderson
committed
Make integration test more readable
Maya Anderson
committed
Comment fix
Maya Anderson
committed
Fix L1 L2 sensitivity parameter documentation
Maya Anderson
committed
junit output format
Maya Anderson
committed
Remove Claude files
Maya Anderson
committed

Commits on Jul 9, 2026

Remove files irrelevant to the PR
Maya Anderson
committed

Commits on Jul 14, 2026

dp_laplace and dp_gaussian now take the original matrix and build a transformation matrix T internally, returning T %*% X with noise fused into a single matrix multiply.
Maya Anderson
committed
Remove 4th and 5th parameters and instead get them in a loop, according to David's suggestion
Maya Anderson
committed
Add dp_set_budget(epsilon, delta) built-in
Maya Anderson
committed
Change DPBudgetAccountant defaults to be constants
Maya Anderson
committed
Cleanup comments
Maya Anderson
committed
Differential Privacy Benchmark
Maya Anderson
committed