README.md

metadata

pretty_name: >-
  The LLM-Attacker: A Unified Red-Team Framework for Privacy Evaluation of Split
  and Distributed LLM Systems
license: apache-2.0

The LLM-Attacker: reproduction and peer-review release

This dataset contains the final paper, frozen experiment evidence, and the source snapshot needed to inspect or reproduce the reported analyses.

The repository is the pre-publication internal-review snapshot. The complete tagged evidence release will become publicly accessible with the paper. Tag v1.7.1 freezes the current manuscript and retained evidence tree.

Final paper

• manuscript-20260930/the-llm-attacker-unified-red-team-framework-for-privacy-evaluation.pdf: current 42-page PDF.
• manuscript-20260930/the-llm-attacker-unified-red-team-framework-for-privacy-evaluation-arxiv-source.zip: self-contained arXiv source package.
• manuscript-20260930/RESULTS-EVIDENCE-MAP.md: claim-to-evidence map for Sections 6 and 7.
• manuscript-20260930/validation/measurement-20261001/: pinned measurement-methodology validation, with 37 passed cases, no failures or skips, raw records and reproduction instructions.
• manuscript-20260930/figure-sources/: editable TikZ and SVG sources for the new framework diagram.
• manuscript-20260930/CHANGES.md: manuscript revision history.
• code/: executable framework and analysis snapshot. Non-functional comments and blank lines were removed; Python abstract syntax is unchanged.

The 30 September revision aligns Paper B's figure and table presentation with Paper A and corrects the manuscript's repository paths, access wording and revision provenance. The v1.4.1 manuscript adds a bounded synthesis of what attack strength is established and a validation matrix accounting for all 64 reviewed modules. It separates reproduced, adapted, real-target qualification, synthetic-only and unmeasured modules, and records each row's model, defence, expected and observed outcome, controls, artifact and status. It also makes the existing heterogeneous evidence across GPT-2 and six Qwen and Mistral models explicit. These runs show framework portability and module qualification, not directly comparable recovery rates. The cumulative evidence is presented as ordinary academic prose rather than as a quotation. No experimental result value changed.

The 1 October v1.5 revision rewrites Results and Scope of the Evidence around the completed experiments. It reports the validated white-box path and the completed ten-setting live-agent qualification directly, removes superseded debugging narratives, and links each major result family to its exact machine-readable evidence file in the frozen v1.4.1 evidence release. The cross-epoch accumulation percentages are no longer used as attack-strength evidence because the underlying report classifies that path as machinery-grade and records a label-alignment defect. No retained validated result changed.

The 1 October v1.6 revision completes a confidence and disclaimer pass across the paper. It states the validated outcomes directly, consolidates scope language, and removes repeated qualifications and superseded failure wording. The experiments, evidence files, tables, figures, and numerical results are unchanged.

The 1 October v1.7 revision expands the framework methodology and research workflows, adds the framework diagram, and names Section 5 “Scoring System and Decision Rules”. Section 6.8 and Table 12 report a fixed selection of 37 measurement-validation cases across seven properties, executed on the reviewed framework commit. All passed, with no failures or skips. The subsection cites Xie et al. (2011) for metamorphic testing and keeps this measurement-conformance study distinct from attack-effectiveness experiments. Existing paragraphs and prior numerical results are preserved. The new supplement accompanies the manuscript; it is not part of the immutable v1.4.1 campaign release.

The 5 October v1.7.1 patch corrects 22 manuscript file references to use Hugging Face file-view URLs (/blob/) instead of folder URLs (/tree/). The evidence remains pinned to v1.4.1; all 20 distinct referenced files were verified through authenticated downloads. Manuscript text and experimental results are unchanged. Dataset access still requires login and approval.

Three revisions have distinct roles:

• manuscript source: GitHub commit 1e9210caf702608e5d0ba09ec3be9293e297d194, with the complete manuscript source in the archive above;
• reviewed 64-module taxonomy: tag submission-candidate-20260922-r2, commit 6c071a68ae5134522d0a4cd3d7d0fa67acabdfa8;
• released code/ base snapshot: commit a3eb53fca5d6395da4c4377e9d8787676c66ae6a, supplemented by experiment-specific source archives and manifests.

The paper's 64-module count is fixed by the reviewed taxonomy tag. The canonical source repository is setloop-io/LLM-Attacker, and this tagged dataset is the peer-review and reproduction artifact.

Main experiment and controls

• v3-campaign/: frozen GPT-2 validation campaign, merged results, shard evidence, and review analyses.
• binary-tier/: binary captures referenced by the main campaign.
• corpora/ and protocols/: fixed inputs, tokenizer data, and validation protocols.
• positive-controls/: undefended and membership-inference controls.

Additional experiments

• domain-shift-v1/: domain-shift measurements.
• long-horizon-v1/: long-horizon measurements.
• clustered-gate-v1/: clustered uncertainty analysis for the gate.
• agentic-live-v1/: live-planner agentic checks.
• all-boundary-whitebox-reconstruction/: validated original-36 and all-boundary white-box packages.
• phase2-raw-evidence/: completed Phase 2 raw evidence.
• bisr-conformance-run-artifacts/: BiSR single-seam conformance run records.
• bisr-conformance-validation/: independently verified BiSR reproduction package.
• bisr-adapted-comparison/: frozen adapted BiSR-versus-joint comparison.
• defence-training-500-step-extension/: six 500-step training trajectories and independent rescoring evidence.
• agentic-nonthinking-decoding/: prospective non-thinking Qwen3 amendment.
• splitllm-reference/: retained reference aggregates.

The dated audit-20260917/ directory is retained as immutable provenance for earlier releases. The current manuscript reports the validated implementation and completed results; superseded manuscript copies and obsolete status documents have been removed from the current branch.

INDEX.md maps paper claims to paths. REPRODUCE.md gives download, verification, and reproduction instructions. MANIFEST.sha256 covers every retained file except .gitattributes and the manifest itself. Frozen evidence packages also retain their internal manifests.