IEEE S&P Artifact Evaluation

Use this once a paper is accepted and the team is deciding whether and how to enter artifact evaluation. At S&P, AE is a post-acceptance opt-in run by a separate committee against three badges (sp2026.ieee-security.org /cfartifacts.html, checked 2026-07-08); the 2027 schedule and any badge changes were unpublished at check time — 待核实 on the current AE pages.

The three badges are three different projects

Pick badges by what an outsider can genuinely re-run. A measurement study of live infrastructure can honestly target Available (data + analysis code) even when Reproduced is impossible; overclaiming a badge wastes the committee's goodwill and your response cycles.

Security artifacts have handling constraints others do not

• Exploit code: release what supports the scientific claim; gate or stub weaponization details when fixes are not universally deployed, and say so in the README. Align with the paper's disclosure timeline — an artifact should not out-run the patch.
• Malware corpora: ship hashes plus a documented retrieval path rather than live binaries; if binaries are necessary, password-protected archives with an explicit handling warning are the community norm.
• Vulnerable targets: containerize the intentionally-vulnerable service so an evaluator cannot accidentally expose it; never require the evaluator to scan or attack third-party infrastructure.
• Datasets with user data: only what the paper's ethics record covers — the AE deposit is public forever.

Packaging for a stranger with a deadline

Evaluators run many artifacts in a fixed window. Optimize for their first thirty minutes:

artifact/
├── README.md            # claims table (below), requirements, total runtime
├── LICENSE
├── Dockerfile           # or VM image reference; pin versions incl. kernel
│                        #   if the attack is kernel-sensitive
├── setup.sh             # one command; no sudo surprises undocumented
├── run_minimal.sh       # <30 min smoke path exercising every component
├── run_full.sh          # regenerates paper numbers; prints ETA up front
└── expected/            # reference outputs + tolerance notes

README claims table:
| Paper claim | Command | Expected output | Time |
| Table 3 attack success | ./run_full.sh t3 | success ≥ 0.9 ± noise note | 2 h |
| Fig 5 overhead curve   | ./run_full.sh f5 | CSV within ±5% of expected/ | 40 m |

Hardware honesty matters more at S&P than most venues: microarchitectural attacks, timing channels, and TEE work are CPU-stepping-sensitive. State the exact tested hardware and what changes off it — "results reproduced on different silicon may differ in " is a credibility line, not a weakness.

Process notes

• Register for AE in the acceptance-notification window; the AE calendar runs in parallel with camera-ready and both compete for the same author-weeks.
• Expect an interactive phase where evaluators file issues; respond with fixes to the artifact, not prose reassurance.
• The badge outcome is independent of the paper's acceptance — a failed AE does not un-accept the paper, but badges print on the published version.
• Keep the deposited DOI version identical to what evaluators approved.

Failure modes specific to this venue

• A "working" exploit that assumed the lab's exact microcode/patch level — pin and document, or downgrade the badge target.
• Disclosure conflict: artifact published before coordinated-disclosure clocks expire.
• The AE README assumes security-tool literacy the committee may not have (e.g., unstated familiarity with a specific fuzzer's corpus format).
• Zenodo deposit made from the wrong branch after camera-ready edits.

Output format

[AE decision] opt in: yes/no — badge targets: Available / Functional / Reproduced
[Rerunnability audit] <what an outsider can actually regenerate>
[Handling constraints] exploit gating / malware packaging / user data: <plan>
[Hardware sensitivity] <exact platform + expected drift off-platform>
[Schedule] AE registration <date 待核实> vs camera-ready <date> — conflict?