diffpriv-pure

Zero-dependency pure-stdlib differential privacy for Python. Laplace, Gaussian, and Geometric mechanisms + a PrivacyBudget sequential accountant — no numpy, no scipy, no build step.

"Ship DP analytics to AWS Lambda in <100 ms cold-start, with zero C extensions."

Quick Start

pip install diffpriv-pure

from diffpriv_pure import laplace_noise, private_sum, private_mean, private_count, PrivacyBudget

# Pure ε-DP: Laplace noise for a sum query
noisy_sum = private_sum(values=[1.2, 3.4, 5.1], bounds=(0.0, 10.0), epsilon=0.5)

# Relaxed (ε, δ)-DP: Gaussian noise for a mean
noisy_mean = private_mean(values=[1.2, 3.4, 5.1], bounds=(0.0, 10.0), epsilon=0.5, delta=1e-6)

# Integer count with geometric (discrete Laplace) noise
noisy_count = private_count(condition_count=42, epsilon=0.5)

# Budget accountant — refuses to overspend ε or δ
budget = PrivacyBudget(total_epsilon=2.0, total_delta=1e-5)
budget.spend(epsilon=0.5, delta=1e-6, query="count_users")
budget.spend(epsilon=0.5, delta=1e-6, query="mean_age")
# budget.spend(epsilon=2.0)  → raises PrivacyBudgetExhausted

CLI

# Laplace mechanism
diffpriv-pure laplace --epsilon 1.0 --sensitivity 1.0 --value 42.0

# Gaussian mechanism
diffpriv-pure gaussian --epsilon 0.5 --delta 1e-6 --sensitivity 1.0 --value 100.0

# Private count (geometric noise)
diffpriv-pure count --epsilon 0.5 --value 1000

# Private mean from a file
diffpriv-pure mean --epsilon 0.5 --bounds 0,100 --input values.txt

JSON output (all commands):

{
  "value": 42.39635838696227,
  "epsilon": 1.0,
  "delta": 0.0,
  "mechanism": "laplace"
}

Public API

⚡ Performance

No heavy dependencies means sub-100 ms cold-start on serverless:

python3 benchmarks/run_benchmark.py

Limitations

• No DP-SGD / federated learning — this is a classical mechanism library, not a deep learning integration.
• No synthetic data generation — no GANs, VAEs, or marginal synthesis.
• No auto-sensitivity estimation — callers must supply bounds or sensitivity manually.
• No async / parallel / vectorised operations — works on list[float], not numpy arrays.
• No advanced composition auto-application — advanced_composition() method is exposed for manual use; basic sequential composition is applied automatically.
• Does not guarantee exact integer output for private_sum / private_mean (these return float).

Non-Goals

This library does not provide:

• DP-SGD or federated learning integration (PyTorch / TensorFlow / JAX)
• Synthetic data or marginal synthesis (GANs, VAEs)
• Auto-sensitivity estimation
• SQL query rewriting
• Privacy-utility tradeoff optimization
• Support for negative δ

Install (clean clone)

git clone https://github.com/prasad-a-abhishek/diffpriv-pure.git
cd diffpriv-pure
pip install -e .

Or install from source with zero dependencies:

pip install --no-deps diffpriv-pure
pip check   # shows no missing dependencies

Requires: Python 3.11+ — stdlib only (math, random, secrets, argparse, json).

Test Suite

pytest --collect-only -q | tail -1
# → 126 tests collected

pytest -v
# → 126 passed in ~0.5s

Coverage map (spec acceptance criteria → tests):

Adversarial QA

For the full adversarial QA report (methodology, surfaces, fuzz harnesses, 300k+ iters, 0 findings), see benchmarks/adversarial/FUZZING_REPORT.md.