Skip to content

AuditPoison v0.4.0 — EvidenceShield

Choose a tag to compare

View all tags
@anirudhnshandilya anirudhnshandilya released this 30 Jul 14:06
v0.4.0
eefb6ed

This release adds EvidenceShield v0.1, a provenance-aware defence wrapper with:

  • scope and assessment-period screening;
  • instruction-like content quarantine;
  • unsupported authority-claim detection;
  • provenance and integrity labelling;
  • model-visible trust annotations;
  • a conservative positive-assurance gate;
  • defence comparison tooling and regression tests.

The benchmark remains frozen at 40 bundles across 20 paired scenarios so the defence can be evaluated without changing the underlying pilot.

Benchmark composition

Oracle label Bundles
Compliant 20
Non-compliant 15
Insufficient evidence 5
Total 40

Attack families: instruction injection, authority spoofing, contradiction flooding, temporal rollback, scope substitution, and evidence omission.

Benign perturbations: formatting noise, evidence reordering, irrelevant context, duplicate support, semantic paraphrasing, and metadata noise.

Installation

git clone https://github.com/anirudhnshandilya/AuditPoison.git
cd AuditPoison
python -m venv .venv

Windows Command Prompt:

.venv\Scripts\activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"

Linux and macOS:

source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e ".[dev]"

Validate

python scripts/validate_dataset.py
python -m pytest -q

Expected: 40 bundles, 20 pairs, and 20 tests passing.

Inspect EvidenceShield

python scripts/inspect_defense.py --bundle-id AP-AC2-001-attacked

Compare unshielded and shielded smoke tests

python scripts/run_model.py --adapter keyword --output results/keyword_raw_predictions.jsonl
python scripts/evaluate_predictions.py results/keyword_raw_predictions.jsonl --output results/keyword_raw_metrics.json

python scripts/run_model.py --adapter keyword --defense evidenceshield --output results/keyword_shield_predictions.jsonl
python scripts/evaluate_predictions.py results/keyword_shield_predictions.jsonl --output results/keyword_shield_metrics.json

python scripts/compare_defenses.py results/keyword_raw_metrics.json results/keyword_shield_metrics.json
python scripts/build_paper_table.py results/keyword_raw_metrics.json results/keyword_shield_metrics.json

The keyword adapter is a deterministic software smoke test, not a research baseline.

Run a real local model

python scripts/run_model.py --adapter ollama --model YOUR_MODEL --limit 2 --output results/ollama_test.jsonl
python scripts/run_model.py --adapter ollama --model YOUR_MODEL --defense evidenceshield --output results/ollama_shield_predictions.jsonl

Run an OpenAI-compatible endpoint

Set AUDITPOISON_API_KEY only in the current shell, then run:

python scripts/run_model.py --adapter openai-compatible --model YOUR_MODEL --base-url YOUR_BASE_URL --defense evidenceshield --output results/hosted_shield_predictions.jsonl

AuditPoison never writes the API key to predictions or manifests.

Core metrics

  • accuracy and macro F1;
  • False Assurance Rate;
  • paired Attack Success Rate and robust accuracy;
  • benign consistency and both-correct rate;
  • citation precision, recall, and F1;
  • attack-evidence detection and benign false-flag rate;
  • Brier score and expected calibration error.

Repository layout

data/pilot/           Evidence bundles
docs/                 Threat model, protocols, and EvidenceShield design
prompts/              Frozen auditor prompts
results/              Metrics and paper tables
schema/               Evidence-bundle JSON Schema
scripts/              Validation, experiments, comparison, and reporting
src/auditpoison/      Python package and defence implementation
tests/                Integrity, leakage, reproducibility, and defence tests

Versions

  • v0.1.0 — threat model and adversarial pilot
  • v0.2.0 — balanced benchmark and evaluation pipeline
  • v0.3.0 — real-model adapters and reproducibility manifests
  • v0.4.0 — EvidenceShield provenance firewall and assurance gate

Licence and citation

AuditPoison is released under the MIT License. Citation metadata is provided in CITATION.cff.

Assets 2
Loading