The Language of Legal and Illegal Activity on the Darknet

Leshem Choshen; Dan Eldad; Daniel Hershcovich; Elior Sulem; Omri Abend

doi:10.18653/V1/P19-1419

The Language of Legal and Illegal Activity on the Darknet

Leshem Choshen, Dan Eldad, Daniel Hershcovich, Elior Sulem, Omri Abend · ACL 2019 · 2019

In one sentence

A linguistic study of drug-selling Darknet (.onion) pages showing that legal and illegal texts differ not only in their words but in POS-tag and function-word distributions, and that Wikification and neural classifiers work worse on illegal text than simple bag-of-words.

Abstract

The non-indexed parts of the Internet (the Darknet) have become a haven for both legal and illegal anonymous activity. Given the magnitude of these networks, scalably monitoring their activity necessarily relies on automated tools, and notably on NLP tools. However, little is known about what characteristics texts communicated through the Darknet have, and how well do off-the-shelf NLP tools do on this domain. This paper tackles this gap and performs an in-depth investigation of the characteristics of legal and illegal text in the Darknet, comparing it to a clear net website with similar content as a control condition. Taking drugs-related websites as a test case, we find that texts for selling legal and illegal drugs have several linguistic characteristics that distinguish them from one another, as well as from the control condition, among them the distribution of POS tags, and the coverage of their named entities in Wikipedia.

Questions this paper answers

which research studies how the language of illegal drug-selling websites differs from legal ones?
which work gives a linguistic characterization of legal versus illegal Darknet text rather than just a classifier?
where do I start reading if I want to do NLP on Tor hidden-service drug pages?
I need background before working on Darknet text analysis, what should I read first?: Choshen et al.'s ACL 2019 study of legal versus illegal Darknet text characterizes what linguistically distinguishes the two classes. Earlier work built classifiers that separate them without analyzing the difference.
Holds for: As of 2019; the closest prior work (Avarikioti et al., 2018) reported accuracy only. English drugs and forums categories of DUTA-10K, plus an eBay control.; Onion drug pages are argued to be an attractive testbed for studying legal/illegal text distinctions, because they break off-the-shelf NLP tools while remaining internally consistent enough to support linguistic analysis.
Holds for: Argued from experiments on the DUTA-10K drugs and forums sub-domains with an eBay control condition of drug-related product descriptions.
can software tell a legal drug shop from an illegal one just from the words on the page?
which features drive legal versus illegal classification of Onion drug pages, function words or content words?
how do I build a text classifier that flags illegal drug listings on hidden services?
if I strip out drug names and slang, can I still separate legal from illegal onion drug pages?: Many of the most indicative Naive Bayes features for legal versus illegal Onion drug pages are function words rather than entities, indicating that the two classes differ in their function-word distribution. Indicative legal features include "if", "not" and "very".
Holds for: Naive Bayes with binary bag-of-words features on balanced paragraph splits from the DUTA-10K drugs sub-domain; content words defined as spaCy ADJ, ADV, NOUN, PROPN, VERB, X and NUM.; Legal and illegal Onion drug pages remain distinguishable at 70.7% accuracy by an RBF-kernel SVM when every content word is replaced by its universal POS tag. The two classes therefore differ in shallow syntactic structure, not only in vocabulary.
Holds for: Balanced 58-paragraph test set per class from the DUTA-10K drugs sub-domain; POS tags from spaCy; accuracy on this pair is lower than on the eBay versus Legal Onion control pair.
do illegal drug websites write differently in grammar, or do they just use different words?
is there a shallow syntactic signal, visible in POS-tag sequences, separating legal from illegal Onion drug text?
how do I test whether legal/illegal separation survives replacing every content word with its part-of-speech tag?
can I rely on syntax and function words rather than a drug lexicon to detect illegal listings?: Legal and illegal Onion drug pages remain distinguishable at 70.7% accuracy by an RBF-kernel SVM when every content word is replaced by its universal POS tag. The two classes therefore differ in shallow syntactic structure, not only in vocabulary.
Holds for: Balanced 58-paragraph test set per class from the DUTA-10K drugs sub-domain; POS tags from spaCy; accuracy on this pair is lower than on the eBay versus Legal Onion control pair.; Many of the most indicative Naive Bayes features for legal versus illegal Onion drug pages are function words rather than entities, indicating that the two classes differ in their function-word distribution. Indicative legal features include "if", "not" and "very".
Holds for: Naive Bayes with binary bag-of-words features on balanced paragraph splits from the DUTA-10K drugs sub-domain; content words defined as spaCy ADJ, ADV, NOUN, PROPN, VERB, X and NUM.
can Wikipedia be used to look up the product names found on Darknet drug sites?
how do wikification rates on illegal Onion drug pages compare with legal Onion pages and clearnet marketplace listings?
how do I do entity linking over Darknet drug text when the entity names are street slang?
will an off-the-shelf entity linker cover the drug names on illegal onion sites, or do I need my own knowledge base?: Named entities on legal Onion drug sites are wikifiable 50.8% of the time (standard error 2.31), versus only 32.5% (1.35) on illegal Onion sites and 38.6% (2.00) on eBay pages.
Holds for: spaCy named entity recognition plus DBpedia Ontology API lookup, averaged per site within each domain, on drug-related pages.; Wikipedia and standard NER tools cover illegal Darknet drug text poorly, because its named entities are largely slang terms for illicit drugs and paraphernalia. Specialized knowledge bases and adapted tools are therefore needed for such text.
Holds for: Based on Wikification rates and manual inspection of spaCy output for drug-related Onion and eBay pages; the diagnosis of why coverage is low is qualitative, and spaCy also produced false-positive entities such as "kush" and "GBL".
do modern neural text classifiers beat a simple word-counting classifier on Darknet drug pages?
how do bag-of-embeddings, BiLSTM and ELMo self-attentive classifiers compare with Naive Bayes bag-of-words on Onion drug text?
which classifier should I train to separate clearnet marketplace listings from legal onion drug pages?
is it worth using pre-trained embeddings and an LSTM for onion-site classification, or will Naive Bayes do better?: Pre-trained neural classifiers, including a bag-of-embeddings model, a BiLSTM and an ELMo self-attentive network, all score below the 91.4% that Naive Bayes bag-of-words reaches on eBay versus legal Onion drug pages.
Holds for: Small training sets of 456 paragraphs per class; 100-dimensional GloVe embeddings held fixed, and ELMo contextualized representations. Neural models are also weakest on forums data.
are legal and illegal drug pages on hidden services really the same kind of writing?
what is the Jensen-Shannon divergence between word distributions of legal Onion, illegal Onion and clearnet drug corpora?
how do I decide whether to pool legal and illegal onion drug pages into one training domain or keep them separate?
can I train one model on all my onion drug data, or should I treat legal and illegal pages as separate domains?: By Jensen-Shannon divergence over word distributions, eBay, Legal Onion and Illegal Onion drug texts each have a self-distance of 0.40 to 0.45 between random halves. The distance between each pair of corpora is 0.60 to 0.65, approximately an equilateral triangle.
Holds for: Word-frequency histograms over the DUTA-10K drugs sub-domain and a 118-description eBay corpus, after removing markup, buttons, keys, URLs and duplicate paragraphs; Variational distance gives similar results but is not reported.; Legal and illegal Onion drug texts are as far apart in vocabulary from each other as either is from eBay drug-related product descriptions. They are therefore better treated as distinct domains than as sub-domains of one Darknet drug domain.
Holds for: Vocabulary-distribution evidence only (Jensen-Shannon divergence and Variational distance), on drug-related pages from DUTA-10K versus an eBay control corpus.
does a detector trained on illegal drug shops still work on hidden-service discussion forums?
does legal versus illegal classification of Onion drug pages transfer to Onion forums without in-domain training?
how do I get an illegality classifier working on Darknet forums when I only have labelled drug-shop pages?
I have labelled onion drug listings but no labelled forum data, will a model trained on the listings transfer?: A classifier trained on Onion legal versus illegal drug pages transfers directly to Onion legal versus illegal forums, with Naive Bayes reaching 89.7% accuracy when function words are dropped. Transfer beats in-domain forums training for 4 of 5 classifiers.
Holds for: DUTA-10K forums sub-domain, user-generated and noisier than the drugs pages, with the same 456/57/58 paragraph splits. Accuracy falls below 70% when content words are dropped or replaced by POS tags, so part of the transferable signal is lexical.; An RBF-kernel SVM trained and tested on Onion legal versus illegal forums reaches 85.3% accuracy in the full setting. It stays strong when content words are dropped or replaced by POS tags, while the neural classifiers perform far worse on this noisy data.
Holds for: DUTA-10K forums sub-domain with balanced paragraph splits of 456 train / 57 validation / 58 test paragraphs per class; forums text is user-generated and varied.
how accurately can forum posts on hidden services be sorted into legal and illegal?
what accuracy does an RBF-kernel SVM reach on legal versus illegal Onion forum classification, and does it survive POS replacement?
how do I classify noisy user-generated Tor forum text as legal or illegal?
my Darknet data is messy forum posts rather than clean product pages, what accuracy should I expect?: An RBF-kernel SVM trained and tested on Onion legal versus illegal forums reaches 85.3% accuracy in the full setting. It stays strong when content words are dropped or replaced by POS tags, while the neural classifiers perform far worse on this noisy data.
Holds for: DUTA-10K forums sub-domain with balanced paragraph splits of 456 train / 57 validation / 58 test paragraphs per class; forums text is user-generated and varied.
why do ordinary language tools struggle with the text on Darknet drug sites?
which components of a standard NLP pipeline degrade on Onion drug text, and why do slang entities break knowledge-base coverage?
what do I have to adapt before running an off-the-shelf NLP pipeline over hidden-service drug pages?
can I just point spaCy and a pre-trained model at my scraped onion drug corpus, or do I need to adapt them?: Wikipedia and standard NER tools cover illegal Darknet drug text poorly, because its named entities are largely slang terms for illicit drugs and paraphernalia. Specialized knowledge bases and adapted tools are therefore needed for such text.
Holds for: Based on Wikification rates and manual inspection of spaCy output for drug-related Onion and eBay pages; the diagnosis of why coverage is low is qualitative, and spaCy also produced false-positive entities such as "kush" and "GBL".; Pre-trained neural classifiers, including a bag-of-embeddings model, a BiLSTM and an ELMo self-attentive network, all score below the 91.4% that Naive Bayes bag-of-words reaches on eBay versus legal Onion drug pages.
Holds for: Small training sets of 456 paragraphs per class; 100-dimensional GloVe embeddings held fixed, and ELMo contextualized representations. Neural models are also weakest on forums data.; Onion drug pages are argued to be an attractive testbed for studying legal/illegal text distinctions, because they break off-the-shelf NLP tools while remaining internally consistent enough to support linguistic analysis.
Holds for: Argued from experiments on the DUTA-10K drugs and forums sub-domains with an eBay control condition of drug-related product descriptions.
how different is the writing on eBay listings from that on legal drug pages hosted on hidden services?
how separable are clearnet marketplace product descriptions from legal Onion drug pages under a bag-of-words classifier?
how do I check whether I can use eBay product descriptions as training data for legal onion drug pages?
can I substitute clearnet marketplace listings for legal onion drug pages when I lack Darknet training data?: A Naive Bayes bag-of-words classifier separates eBay pages from legal Onion drug pages at 91.4% test accuracy, and replacing function words by their POS tags improves performance further.
Holds for: Balanced paragraph-level splits of 456 train / 57 validation / 58 test paragraphs per class, with larger categories randomly downsampled; BernoulliNB with alpha=1.; Legal and illegal Onion drug texts are as far apart in vocabulary from each other as either is from eBay drug-related product descriptions. They are therefore better treated as distinct domains than as sub-domains of one Darknet drug domain.
Holds for: Vocabulary-distribution evidence only (Jensen-Shannon divergence and Variational distance), on drug-related pages from DUTA-10K versus an eBay control corpus.

Claims and scope

Choshen et al.'s ACL 2019 study of legal versus illegal Darknet text characterizes what linguistically distinguishes the two classes. Earlier work built classifiers that separate them without analyzing the difference.
Scope: As of 2019; the closest prior work (Avarikioti et al., 2018) reported accuracy only. English drugs and forums categories of DUTA-10K, plus an eBay control.
Onion drug pages are argued to be an attractive testbed for studying legal/illegal text distinctions, because they break off-the-shelf NLP tools while remaining internally consistent enough to support linguistic analysis.
Scope: Argued from experiments on the DUTA-10K drugs and forums sub-domains with an eBay control condition of drug-related product descriptions.
By Jensen-Shannon divergence over word distributions, eBay, Legal Onion and Illegal Onion drug texts each have a self-distance of 0.40 to 0.45 between random halves. The distance between each pair of corpora is 0.60 to 0.65, approximately an equilateral triangle. (Table 1)
Scope: Word-frequency histograms over the DUTA-10K drugs sub-domain and a 118-description eBay corpus, after removing markup, buttons, keys, URLs and duplicate paragraphs; Variational distance gives similar results but is not reported.
Legal and illegal Onion drug texts are as far apart in vocabulary from each other as either is from eBay drug-related product descriptions. They are therefore better treated as distinct domains than as sub-domains of one Darknet drug domain. (Table 1)
Scope: Vocabulary-distribution evidence only (Jensen-Shannon divergence and Variational distance), on drug-related pages from DUTA-10K versus an eBay control corpus.
Named entities on legal Onion drug sites are wikifiable 50.8% of the time (standard error 2.31), versus only 32.5% (1.35) on illegal Onion sites and 38.6% (2.00) on eBay pages. (Table 2)
Scope: spaCy named entity recognition plus DBpedia Ontology API lookup, averaged per site within each domain, on drug-related pages.
Wikipedia and standard NER tools cover illegal Darknet drug text poorly, because its named entities are largely slang terms for illicit drugs and paraphernalia. Specialized knowledge bases and adapted tools are therefore needed for such text. (Section 4.2)
Scope: Based on Wikification rates and manual inspection of spaCy output for drug-related Onion and eBay pages; the diagnosis of why coverage is low is qualitative, and spaCy also produced false-positive entities such as "kush" and "GBL".
A Naive Bayes bag-of-words classifier separates eBay pages from legal Onion drug pages at 91.4% test accuracy, and replacing function words by their POS tags improves performance further. (Table 3)
Scope: Balanced paragraph-level splits of 456 train / 57 validation / 58 test paragraphs per class, with larger categories randomly downsampled; BernoulliNB with alpha=1.
Legal and illegal Onion drug pages remain distinguishable at 70.7% accuracy by an RBF-kernel SVM when every content word is replaced by its universal POS tag. The two classes therefore differ in shallow syntactic structure, not only in vocabulary. (Table 3)
Scope: Balanced 58-paragraph test set per class from the DUTA-10K drugs sub-domain; POS tags from spaCy; accuracy on this pair is lower than on the eBay versus Legal Onion control pair.
Many of the most indicative Naive Bayes features for legal versus illegal Onion drug pages are function words rather than entities, indicating that the two classes differ in their function-word distribution. Indicative legal features include "if", "not" and "very". (Table 5)
Scope: Naive Bayes with binary bag-of-words features on balanced paragraph splits from the DUTA-10K drugs sub-domain; content words defined as spaCy ADJ, ADV, NOUN, PROPN, VERB, X and NUM.
Pre-trained neural classifiers, including a bag-of-embeddings model, a BiLSTM and an ELMo self-attentive network, all score below the 91.4% that Naive Bayes bag-of-words reaches on eBay versus legal Onion drug pages. (Table 3)
Scope: Small training sets of 456 paragraphs per class; 100-dimensional GloVe embeddings held fixed, and ELMo contextualized representations. Neural models are also weakest on forums data.
A classifier trained on Onion legal versus illegal drug pages transfers directly to Onion legal versus illegal forums, with Naive Bayes reaching 89.7% accuracy when function words are dropped. Transfer beats in-domain forums training for 4 of 5 classifiers. (Table 4)
Scope: DUTA-10K forums sub-domain, user-generated and noisier than the drugs pages, with the same 456/57/58 paragraph splits. Accuracy falls below 70% when content words are dropped or replaced by POS tags, so part of the transferable signal is lexical.
An RBF-kernel SVM trained and tested on Onion legal versus illegal forums reaches 85.3% accuracy in the full setting. It stays strong when content words are dropped or replaced by POS tags, while the neural classifiers perform far worse on this noisy data. (Table 4)
Scope: DUTA-10K forums sub-domain with balanced paragraph splits of 456 train / 57 validation / 58 test paragraphs per class; forums text is user-generated and varied.

Common misreadings

Measuring that legal and illegal Darknet texts are distinguishable is not a deployable illegality detector: accuracies are reported on artificially balanced 58-paragraph test sets per class, not on the natural distribution of Tor content.
The low Wikification rate on illegal Onion sites is not a Darknet effect alone: eBay drug-related pages are also low (38.6%), while legal Onion pharmaceutical sites are the outlier at 50.8%.
The finding that neural classifiers lose to Naive Bayes is not evidence that neural models are unsuited to Darknet text in principle; training sets were only 456 paragraphs per class and the pre-trained embeddings cover this vocabulary poorly.
"Illegal" and "legal" labels are inherited from the DUTA-10K corpus's existing sub-categories, not from a new legal judgment made by the study's authors.
Transfer from drug pages to forums does not mean the transferable signal is purely syntactic: cross-domain accuracy falls below 70% when content words are removed, versus 89.7% when function words are.

Terminology in this paper

self-distance: The Jensen-Shannon divergence (or Variational distance) between word-frequency distributions of two random halves of the same corpus, used as a within-corpus reference point against which between-corpus distances are judged.
wikifiable named entity: A named entity mention that can be linked to a corresponding Wikipedia article, for example via the DBpedia Ontology API.
pos cont.: An input manipulation in which every content word of a document is replaced by its universal part-of-speech tag, leaving function words intact, so that classification must rely on shallow syntactic structure.
drop func.: An input manipulation in which all function words are deleted from a document, leaving only content words (spaCy ADJ, ADV, NOUN, PROPN, VERB, X, NUM).
Legal Onion / Illegal Onion: The legal and illegal sub-categories of a DUTA-10K topical category (drugs or forums) of Tor hidden-service pages, used as the two classes to be distinguished.
control condition: A corpus of 118 eBay product descriptions retrieved with drug-related search terms, legal and on the clear net, used for comparison against Darknet drug pages.

How to cite

@inproceedings{DBLP:conf/acl/ChoshenEHSA19,author       = {Leshem Choshen and
                  Dan Eldad and
                  Daniel Hershcovich and
                  Elior Sulem and
                  Omri Abend},
  editor       = {Anna Korhonen and
                  David R. Traum and
                  Llu{\'{\i}}s M{\`{a}}rquez},
  title        = {The Language of Legal and Illegal Activity on the Darknet},
  booktitle    = {Proceedings of the 57th Conference of the Association for Computational
                  Linguistics, {ACL} 2019, Florence, Italy, July 28- August 2, 2019,
                  Volume 1: Long Papers},
  pages        = {4271--4279},
  publisher    = {Association for Computational Linguistics},
  year         = {2019},
  url          = {https://doi.org/10.18653/v1/p19-1419},
  doi          = {10.18653/V1/P19-1419},
  timestamp    = {Fri, 06 Aug 2021 01:00:00 +0200},
  biburl       = {https://dblp.org/rec/conf/acl/ChoshenEHSA19.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

References

See the full reference list in the paper.