Iterative leak-surface peeling

A falsifiability-anchored methodology for ML detection of blockchain validator infrastructure attacks.

Author: Simon Morley, NullRabbit

Status: working draft (2026-05-06). Updates land directly in paper.md; the v1 preprint ships when arXiv submission lands.

---

Abstract

Public-internet-reachable blockchain validator infrastructure is an inviting target for asymmetric resource-consumption attacks; validator-DoS findings have accumulated steadily across Sui, Solana, Ethereum, Aptos, and Cosmos-family chains through 2024–2026 coordinated-disclosure pipelines. Detection that survives the cleartext / TLS-fronted boundary, the cross-chain transfer, and the lab / mainnet fidelity tier is non-optional for operators standing up monitoring on this threat surface. Yet ML-based detection of these attacks routinely ships with median ROC-AUC at or near 1.000 across cross-validation folds, which on audit prove artefactual rather than reflecting underlying network behaviour [sommer2010outside, arp2022dosdonts]. Our V1 trainer's first TrainReport against an initial 1,092-bundle Sui validator-DoS corpus is a fresh data point in this pattern: ROC = 1.000 across all 17 leave-one-primitive-out folds, driven by random-Gaussian-feature ROC = 1.000 (LOPO benign-holdout contamination) and resp.count ROC = 1.000 (capture-pipeline co-linearity).

This paper presents iterative leak-surface peeling: a methodology that pre-registers numerical thresholds, structural predictions, and outcome-band composition rules before headline numbers land, then surfaces and closes its own failure modes through multi-layer falsifiability discriminators and audit-driven iteration. Across V1 → V7-narrow on a 2,103-bundle Sui+Solana corpus, the methodology has surfaced and closed eight distinct leak surfaces in sequence. Two open pillars substantiate the methodology: Bundle v1, a multi-modal capture format with controlled-vocabulary provenance fields, demonstrated through three landed schema-additive extensions; and a chain-agnostic family taxonomy of ten attack families plus benign that classifies post-cycle attacks without extension.

The technical centerpiece is the V7-narrow multi-mechanism finding: cross-chain mechanism non-transfer in this domain is feature-localised, not whole-feature-space. The headline outcome at the V7-narrow evaluation gate is Joint C (cross-chain mechanism claim at the rate-invariant 13-feature manifest layer falsified at lab fidelity). Subsequent Step-11 V1 + V8 retrain cycles deliver Joint A on the cipher-agnostic claim: 100% retention vs the pre-registered Step-11 V1 cleartext baseline across both chains in both LOPO regimes (Step-11 V1 retrain), and 124–191% retention post-trim (Step-11 V8 retrain) once the V7-narrow §SE3-affected pcap.mean_packet_size distribution-mismatch source is removed from the manifest. The encryption boundary itself does not introduce additional accuracy degradation. The methodology contributions persist independently of any single evaluation-gate outcome; the v2 trigger becomes V9-or-later cycles along orthogonal axes (cipher-suite variation, mTLS, additional chains).

Repo layout

• paper.md - the working preprint, canonical source. Updates land here.
• bibliography.bib - references (BibTeX).
• LICENSE-TEXT.md - CC-BY-4.0 (paper text + figures + tables).
• LICENSE-CODE.md - MIT (any scripts / build tooling).

Companion artefacts

• nr-bundle-spec v0.1.0 - open format specification of Bundle v1 (github.com/NullRabbitLabs/nr-bundle-spec, MIT-licensed; private repo until 2026-06-05 pending coordinated-disclosure window closure on Sui F10/F14 + Solana F10, public after). JSON Schema + Python and Rust reference parsers + 5 example bundles + CI for schema regen and cross-language consistency.
• nr-bundles-public (forthcoming, HuggingFace) - curated 20–50 sample bundles spanning multiple primitives + chains.

Sister paper

Earned Autonomy (Zenodo DOI 10.5281/zenodo.18406828) - architectural-layer companion for production deployment. This paper is the data-layer companion; standalone-readable, cross-references where load-bearing.

Citing

A formal arXiv preprint identifier lands at v1 ship. Until then, cite this working draft as:

@misc{morley2026leaksurfacepeeling,
  author = {Morley, Simon},
  title  = {Iterative leak-surface peeling: a falsifiability-anchored
            methodology for {ML} detection of blockchain validator
            infrastructure attacks},
  year   = {2026},
  note   = {Working draft, 2026-05-06.
            \url{https://github.com/NullRabbitLabs/nr-substrate-paper}}
}

License

Dual-licensed:

• Paper text + figures + tables: CC-BY-4.0 (LICENSE-TEXT.md).
• Code, scripts, build tooling: MIT (LICENSE-CODE.md).