Setloop/The-LLM-Attacker-Unified-Red-Team-Framework-for-Privacy-Evaluation-of-Split-and-Distributed-LLMs

The LLM-Attacker — v1.0

In plain language

Most "this defence works" claims can't be compared — every study measured something different. We built the ruler. LLM-Attacker is an open framework that forces privacy evaluations to say exactly what the attacker saw, what it knew, and what it was allowed to spend. Attacks replay the same recorded traffic, predictions are kept for independent re-scoring, and no defence can pass without a linked undefended reference. Running it on real systems showed that changing a single measurement choice can flip the conclusion — including the gradient-leakage findings of the companion studies (arXiv:2609.04382, arXiv:2610.04128).

This release contains the combined concise paper and full supplement, the retained experimental evidence, and the executable source snapshots needed to inspect or reproduce the reported analyses. The release tag is v1.0.

Paper and source

• Combined manuscript: the concise main text followed by S-numbered supplementary material and one bibliography.
• arXiv source ZIP: one top-level main.tex, the included supplementary source, precompiled bibliography and all five required figures. The extracted bundle compiles with pdfLaTeX.
• Editable paper sources.
• Source repository: maintained framework package 1.0.0.

The authors are Georgios Politis and Evangelos Pappas; the title layout includes Setloop.io between the author blocks. The manuscript is prepared for arXiv submission; it is not presented as accepted conference work and no arXiv identifier has yet been assigned. The chosen arXiv distribution licence is arXiv's perpetual, non-exclusive licence.

Before arXiv submission, confirm prior AI systems and their tasks and the status/overlap of related papers, then complete the acknowledgement. EuroS&P also needs verified anonymous reviewer access to code and complete evidence. These submission actions remain open; this repository release does not complete them.

Scope of the evidence

The reviewed paper catalogue covers 64 modules. The maintained GitHub package registers 65, including secure-compute-audit, which is outside that historical qualification matrix. Implementation qualification and synthetic checks are distinct from real-target privacy measurements. The strongest controlled reconstruction comparisons concern GPT-2 124M under the declared corpora, lengths, observations, knowledge and budgets; additional Qwen and Mistral campaigns establish their own protocol-specific outcomes.

The release retains the original experimental files and hashes. Historical source identifiers and experimental names inside frozen records remain provenance labels, not active Git navigation targets. All current navigation uses v1.0; the source snapshots below preserve reproduction independently of the former repository history.

Retained artifacts

Path	Contents
v3-campaign/	Main GPT-2 measurements, predictions, journals and paired analyses
binary-tier/	Captures referenced by the campaign
corpora/, protocols/, positive-controls/	Pinned inputs, evaluation protocols and controls
domain-shift-v1/, long-horizon-v1/, clustered-gate-v1/	Corpus, training-horizon and uncertainty analyses
all-boundary-whitebox-reconstruction/, phase2-raw-evidence/	Validated white-box/compression studies and retained raw evidence
bisr-conformance-validation/, bisr-conformance-run-artifacts/, bisr-adapted-comparison/	Published-implementation checks and adapted comparison
defence-training-500-step-extension/	Frozen training extension and rescoring
agentic-live-v1/, agentic-nonthinking-decoding/	Distinct planner protocols and negative outcomes
code/, source-snapshots/	Retained base code plus scoped reviewed/base executable archives
validation/measurement-20261001/	Frozen 37-case measurement validation, exact source and Git-independent runner

The validation package's source-snapshot.tar.gz matches all 123 retained source-file hashes. The original selection, results and manifest remain unchanged. reproduce_from_snapshot.py verifies the archive before running the frozen selection. See REPRODUCE.md for download, integrity, rescoring and reproduction commands. These conformance checks are not independent empirical attack results.

Access and licences

This dataset has a publicly visible listing. Downloading protected files currently requires a Hugging Face account and manual approval; the linked source repository currently requires authorised GitHub access. This named release does not provide anonymous conference-review access.

Framework code and release packaging carry MIT licence metadata and the supplied LICENSE. To contribute attack modules, defences or fixes, see the source repository's CONTRIBUTING.md; contributions are licensed under the MIT licence. Third-party models, corpora and dependencies retain their own terms. The paper's arXiv distribution choice is separate. Captures and recovered text should be handled under the applicable data permissions. The clinical-style text is synthetic and does not contain patient records.

Integrity and use

MANIFEST.sha256 lists SHA-256 digests for the retained release files, excluding .gitattributes and the manifest itself. Package-specific manifests are preserved unchanged. LFS-backed files are identified by their immutable SHA-256 object identifiers; RELEASE.json records the release's verification scope. INDEX.md maps the principal evidence families.

Use the framework only for systems and data for which you have authorisation. Results apply to the declared protocol; untested attacks remain untested, and a scorecard pass is not universal privacy assurance.