Artifact: SoK: Colluding Adversaries in Machine Learning Pipelines

USENIX Security 2026 artifact for the paper SoK: Colluding Adversaries in Machine Learning Pipelines. Reproduces three empirical studies on unintended interactions between independent attacks against machine-learning models:

• Part A (paper §5.2, Table 5): data-poisoning vs. model extraction on CIFAR-10/100 — negative collusion.
• Part B (paper §5.3, Table 6): model extraction vs. distribution inference on CelebA/UTKFace — positive collusion.
• Part C (paper §5.4): data reconstruction (Geiping NeurIPS 2020 gradient inversion) vs. membership, attribute, and distribution inference on CIFAR-10 and CelebA/UTKFace — positive collusion.

Targets all three USENIX Security 2026 badges (Artifacts Available, Functional, Results Reproducible). Single locked Python environment (pyproject.toml + uv.lock) pinning amuletml==0.5.1. One setup.sh covers all three parts; each part ships a smoke test (~3-10 min) and a full reproduction script. See README.md for claims, requirements, time budgets, and reference numbers.