Research
Extracting and Verifying Illicit Bitcoin Addresses from Underground Forum Discussions
Most public datasets of "illicit" Bitcoin addresses come from community abuse reports, blockchain clustering heuristics, single-incident collections, or proprietary intelligence vendors, and the…

- arXiv
- 2608.13930
- Published
- 2026-08-14
- Authors
- Abdoul Nasser Hassane Amadou, Arnaud Legout, Imane Fouad, Konstantin Avrachenkov, Anas Motii
AI summary
Overview
Research area: Blockchain forensics and cybercrime measurement — specifically, building labeled datasets of illicit cryptocurrency addresses and validating them against underground-forum evidence.
Technical level: Intermediate. The paper is readable without deep cryptography background, but it assumes familiarity with Bitcoin address formats, on-chain transaction data, and the general problem of label provenance in machine-learning datasets.
Scope: The paper describes a reproducible three-stage pipeline (regex extraction, on-chain validation, LLM screening plus expert review) applied to the CrimeBB HackForums archive and releases a dataset of 2,438 manually verified illicit Bitcoin addresses spanning 2010–2024 across twelve cybercrime categories.
What This Paper Is About
Most public datasets of "illicit" Bitcoin addresses come from community abuse reports, blockchain clustering heuristics, single-incident collections, or proprietary intelligence vendors, and the steps used to build them are rarely reproducible. That makes it hard to know whether a label is actually backed by evidence that the address was involved in crime, and hard to compare detection methods fairly across datasets. The authors address this by mining an underground cybercrime forum directly, so that every labeled address comes with the forum discussion that justifies the label, plus confirmation that the address exists and has transacted on-chain.
Key Contributions
-
A reproducible labeling pipeline for underground forums. The pipeline chains on-chain validation, zero-shot LLM screening, and manual expert verification, and the full pipeline is released alongside the data rather than described only in prose.
-
A released dataset of 2,438 manually verified illicit Bitcoin addresses drawn from fifteen years of HackForums activity, each supported by both forum context and on-chain transaction history, with temporal metadata and cybercrime category annotations. The dataset is available at https://github.com/osteoner/Hackforums-dataset.
-
A transparent label-provenance model. Unlike datasets flagged as relying on proprietary intelligence (the paper's comparison table marks Elliptic, Elliptic++, and Elliptic2 as having no label verifiability), every label here is derivable from public forum text plus public blockchain data.
-
A quantified characterization of the address corpus, including on-chain volume comparisons between illicit and unknown addresses, longitudinal trends from 2010 to 2024, corroboration profiles by crime category, and an estimated recall figure for the LLM screening stage.
Main Findings
-
Scale of the source archive: the HackForums snapshot (dated 2024-09-04) contains 42,474,325 posts across 212 subforums, 4,148,196 threads, 348,340 contracts, and 716,058 users, with an average post length of 27.61 words and average thread title length of 5.29 words.
-
Extraction funnel: regular expressions over five cryptocurrencies (BTC, ETH, LTC, DASH, XMR) produced 146,875 candidate strings; on-chain validation (API queries on 11–12 May 2025) confirmed 27,131 real addresses. Bitcoin dominated with 23,882 addresses (88.0%), followed by Litecoin (3,084), Ethereum (155), and Dash (10). Monero addresses could not be validated at all because their transaction history is hidden by protocol-level privacy features.
-
LLM screening reduced the review burden by roughly an order of magnitude: Qwen3-30B-A3B narrowed 23,882 validated Bitcoin addresses down to 2,606 candidates for human review.
-
Expert verification outcome: of the 2,606 candidates, 2,438 were confirmed illicit, 98 were reclassified as licit, and 70 were excluded for insufficient evidence — an LLM screening precision of 93.6%.
-
Category distribution: Fraud is the largest category at 841 addresses (34.50%), followed by Malware at 329 (13.49%), Investment schemes at 311 (12.76%), Services (hacking-for-hire, DDoS, phishing) at 296 (12.14%), Other at 335 (13.74%), Laundering at 161 (6.60%), Account Selling at 82 (3.36%), Membership at 41 (1.68%), Abuse at 15 (0.62%), Card Dumps at 13 (0.53%), Drugs at 11 (0.45%), and Weapons at 3 (0.12%). The LLM did not confidently assign any address to the Counterfeit category.
-
Illicit addresses move more money: median BTC received is 0.1737 for illicit addresses versus 0.0180 for unknown addresses (roughly ten times larger), and median BTC sent is 0.1616 versus 0.0172. Median transaction count is 7 for illicit addresses versus 4 for unknown ones.
-
Sharing peaked in 2014 and collapsed after 2021: address sharing rose sharply from late 2013, peaked in 2014 — coinciding with early crypto-ransomware such as CryptoLocker, active from September 2013 to early 2014 — stayed elevated through 2017, and then declined. Newly shared valid addresses drop to nearly zero after 2022, which the authors attribute to migration toward private messaging platforms such as Telegram rather than a decline in illicit activity itself.
-
Corroboration varies by crime type: across the seven named categories (2,061 addresses), 837 addresses (40.6%) appear in multiple forum contexts — 512 self-advertised across multiple threads by the same user and 325 reported by multiple independent users. Fraud has the most victim-reported addresses (185), while Investment (156 self-advertised, 50.2%) and Malware (134 self-advertised, 40.7%) show the highest self-advertising rates.
-
Screening recall is imperfect but measured: a manual review of 200 randomly sampled addresses from the 21,276 that the LLM did not flag found 8 illicit addresses, a 4.0% false-negative rate. Extrapolated, this implies roughly 851 missed illicit addresses and an estimated screening recall of 74.1%. The missed cases split into four "firm" examples with explicit transactional text (e.g., a payment destination for a malware-spreading method sold via autobuy, a one-month MailMurder subscription receipt, payment for the TOPHER crypter, and funding for an Orcus RAT variant) and four "arguable" ones where illicit intent must be inferred from thread titles, co-advertisements, or signature blocks.
-
The LLM misses implicit context, not explicit context: the failure mode is addresses whose illicit nature is embedded in forum conventions such as signature blocks and dispute threads, rather than stated in the post itself.
-
Annotation limits are disclosed: verification was done by a single domain expert, so no inter-rater reliability score (e.g., Cohen's Kappa) can be reported, and the expert validated only the binary illicit/licit decision — the sub-category was carried over from the LLM's zero-shot output and may contain misassignments.
Methodology in Plain English
The authors start from an existing academic archive of underground forums (CrimeBB) and take a single HackForums snapshot. They sweep every post and every recorded trade contract with pattern-matching rules tuned to five cryptocurrencies, deliberately keeping everything that looks like an address even if it is probably a hash or a code fragment, to avoid missing real ones.
That produces a lot of junk, so the second stage asks the actual blockchains. Each candidate is looked up through public blockchain APIs, and only strings that resolve to a real address with at least one confirmed transaction survive. This filters out typos, hashes, and cross-chain collisions — for example, legacy Bitcoin P2SH addresses and legacy Litecoin addresses both start with "3".
Reviewing tens of thousands of addresses by hand is impractical, so the third stage uses a language model as a filter, not an oracle. For each address the authors gather up to 20 forum posts mentioning it, prioritizing the earliest appearances, and trim each post to a ±200-character window around the address to keep prompts short. A locally served Qwen3-30B-A3B model is asked, zero-shot at temperature 0.3, to assign exactly one of twelve cybercrime categories (or "Unclear"), returning JSON. Anything that lands in a crime category becomes a candidate; "Unclear" addresses are dropped.
Finally, a domain expert reads every post associated with each candidate, along with the surrounding thread, and assigns illicit, licit, or excluded. Only this human step determines the final label; the model only decides what gets looked at. The authors then audit the model's blind spots by hand-checking a random sample of addresses the model had let through as non-illicit.
Why This Matters
Impact on research. The paper attacks a problem that is usually glossed over: where labels come from. Because most large Bitcoin datasets are built from proprietary intelligence or from clustering heuristics that can propagate errors beyond originally verified addresses, benchmark comparisons between detection methods are hard to interpret. A dataset where each label traces back to a specific forum discussion plus a specific on-chain transaction gives other researchers something to audit and a fairer basis for evaluating classifiers. The explicit recall estimate (~74.1%) and precision figure (93.6%) also let downstream users reason about the label noise they are inheriting.
Real-world applications:
- Law enforcement triage: investigators can prioritize addresses that already carry documented criminal context rather than starting from raw blockchain data.
- Exchange and payment compliance: compliance teams can cross-reference customer deposit addresses against evidence-backed labels when assessing suspicious activity.
- Threat intelligence on criminal services: the category breakdown and self-advertising patterns expose which crime types (Investment schemes, Malware) rely on repeated public promotion of a single payment address.
- Longitudinal cybercrime research: the 2010–2024 timeline documents the shift of criminal commerce from open forums to private messaging, which is useful context for anyone interpreting declining forum-based signals.
Industry relevance. Blockchain analytics vendors, ransomware negotiators, and anti-fraud teams all depend on labeled address lists. This work demonstrates a lower-cost, auditable alternative to buying proprietary feeds, and the precision/recall tradeoffs it reports are directly relevant to anyone deploying an LLM as a first-pass filter in a compliance or investigation workflow.
Future Directions
- Improving recall on implicit and obfuscated context. The 74.1% estimated recall reflects a zero-shot, English-dominant model. Fine-tuning on underground forum dialects, signature blocks, and dispute-thread conventions is the obvious next step, as the authors themselves suggest.
- Multi-annotator verification. With a single expert and no inter-rater reliability score, the ceiling on label confidence is unclear. Independent double-annotation would let future work report agreement statistics and quantify category-level error.
- Re-examining the LLM-assigned sub-categories. The expert verified only illicit versus licit, so the twelve-category taxonomy is inherited unverified from the model. A human-verified taxonomy, or a measured error rate per category, would strengthen the annotation layer considerably.
- Accounting for the migration off public forums. Since the pipeline sees almost no new addresses after 2022, extending evidence collection to private messaging platforms, or pairing forum data with other public evidence sources, is needed to keep this kind of dataset current.
Target Audience
Blockchain analytics researchers and graduate students working on illicit-address detection; dataset curators and machine-learning practitioners who care about label provenance and benchmark quality; law enforcement and exchange compliance analysts who use address labels operationally; and cybercrime researchers studying underground forum economies and the shift from public forums to private channels. Readers looking for novel detection algorithms will not find them here — the value is in the data-construction methodology and the released artifact.
Authors’ abstract
Existing labeled Bitcoin datasets are largely derived from community-reported abuse, blockchain heuristics, incident-specific collections, or proprietary labeling processes. Their construction methods are rarely publicly reproducible and often provide limited evidence that an address was directly involved in illicit activity. We present a reproducible pipeline for constructing evidence-backed Bitcoin labels from HackForums, an underground cybercrime forum with fifteen years…
The full abstract on arXiv