DiffAudit Research

Privacy risk auditing for diffusion models.

English | 中文

DiffAudit Platform · Documentation · Getting Started · Data And Assets · Experiment Status · Security

---

DiffAudit Research implements membership inference attacks and defenses for diffusion models. It covers three attacker knowledge levels — black-box, gray-box, and white-box — and tracks each method from paper review through reproducible experiments.

This repository is part of the DiffAudit system. It focuses on research code and experiment tracking; the product UI lives in DiffAudit Platform, and job scheduling lives in Runtime-Server.

Each research direction is tracked through a consistent set of stages:

Track	Role
Paper baselines	Reproduce or adapt known attack and defense methods as reference points.
New method exploration	Test new ideas with explicit hypotheses and conclusions.
Verified results	Only reviewed, reproducible experiments are promoted to this level.

How It Fits

flowchart LR
  Papers["Papers and published methods"] --> Research["DiffAudit Research"]
  Ideas["New research questions"] --> Research
  Research --> Results["Experiment results and status tracking"]
  Results --> Runtime["Runtime job execution"]
  Results --> Platform["DiffAudit Platform"]
  Platform --> Reports["Audit reports and exports"]

Loading

Quick Start

git clone https://github.com/DeliciousBuding/DiffAudit-Research.git
cd DiffAudit-Research
conda env create -f environment.yml
conda activate diffaudit-research
python scripts/bootstrap_research_env.py --install
python scripts/verify_env.py
python -m diffaudit --help

Large datasets and model weights are not stored in Git. See docs/assets-and-storage/data-and-assets-handoff.md for how to set up local data paths.

Documentation

Need	Start here
New contributor setup	docs/start-here/getting-started.md
Environment setup	docs/start-here/teammate-setup.md
Datasets and model weights	docs/assets-and-storage/data-and-assets-handoff.md
CLI commands	docs/start-here/command-reference.md
Experiment status	docs/evidence/reproduction-status.md
Platform integration	docs/product-bridge/README.md
Repository structure	docs/start-here/repo-map.md
Full documentation index	docs/README.md

Repository Layout

Path	What's inside
src/diffaudit/	Python package and CLI — attack methods, defense methods, metrics, utilities.
configs/	Experiment configs and local path templates.
tests/	Test suite.
scripts/	Setup, validation, and experiment scripts.
docs/	Contributor guide, experiment status, platform integration docs.
workspaces/	Current research state for each direction.
legacy/	Archived experiment notes and history.
external/	Upstream code clones (git-ignored).
third_party/	Vendored upstream code subsets with license notices.

Experiment Tracking

Each research direction has a tracking status indicating its maturity:

Status	Meaning
research-ready	Paper, upstream code, and data requirements reviewed.
code-ready	Commands, configs, and tests exist in this repository.
asset-ready	Required datasets or model weights are available locally.
evidence-ready	A reviewed experiment summary exists.
benchmark-ready	Paper-level benchmarks can be reproduced.

Smoke tests and dry runs are engineering checks, not benchmark results. Negative results are kept to avoid repeating failed experiments.

Citation and License

To cite DiffAudit Research, use CITATION.cff. Upstream papers, datasets, and third-party code should be cited under their own terms.

Source code, configs, tests, scripts, and original documentation are licensed under the Apache License 2.0. See docs/governance/licensing.md and NOTICE for third-party license details.