A neural-network security-analysis framework for offensive and defensive AI security operations, supporting two CCS 2026 submissions that share this codebase but evaluate disjoint claims.
This top-level README is an index. Reviewers should follow the paper-specific README for the submission they are reviewing:
| If you are reviewing… | Follow… | One-command reproduction |
|---|---|---|
| The offensive paper "NeurInSpectre: An Offensive Framework for Breaking Gradient Obfuscation in AI Safety Systems via Spectral-Volterra-Krylov Analysis" |
README_OFFSEC.md → AE.md |
bash scripts/reproduce_table8.sh |
| The detection paper "NeurInSpectre: A Three-Layer Mathematical Framework for Gradient Obfuscation Detection in Adversarial Machine Learning" |
README_DETECTION.md → QUICKSTART_CCS.md |
bash scripts/reproduce_detection.sh |
The two reproduction paths are fully independent — outputs are
written to non-overlapping subdirectories (results/table8_run_v2/ +
results/table5_rigor_production/ for the offensive paper;
results/detection/ for the detection paper). You only need to run the
quickstart for the paper you are reviewing.
- Code:
neurinspectre/(~77 000 non-blank lines of GPU-accelerated Python). Single Python package shared by both papers; each paper exercises a different subset. - Scripts:
scripts/reproduce_*.sh(per-paper harnesses) + per-module CLIs. - Models:
models/(TorchScript artifacts pinned by SHA-256 in*.meta.jsonsidecars). The detection paper additionally pulls the Carmon2019Unlabeled checkpoint viascripts/download_carmon2019.py. - As-run artifacts:
results/(single-seed outputs that back the paper tables; the offensive paper's main matrix isresults/table8_run_v2/, the detection paper's isresults/detection/). - Audit / claim trace:
DRAFT_CODE_AUDIT.md(exhaustive claim → code → command → artifact mapping for the offensive paper).
python3.10 -m venv .venv-neurinspectre
source .venv-neurinspectre/bin/activate
pip install -U pip
pip install -e ".[dev]"
pip install -r requirements-frozen.txt
neurinspectre doctorAfter install, jump to whichever paper-specific README applies (table at top).
# Offensive paper smoke test (~5 minutes; exercises the full Table 8 pipeline)
neurinspectre table2-smoke --output-dir results/smoke
# Detection paper smoke test (~5 seconds; Tables 1–4 synthetic only)
python scripts/run_synthetic_experiments.py --output-dir results/smokeMIT (see LICENSE).
For double-blind review questions, please use the OpenReview / HotCRP submission system. Post-acceptance contact details will be added here.