A Broad Evaluation of the Tor English Content Ecosystem

TL;DR

This study provides a comprehensive analysis of the English Tor ecosystem by crawling over 1 million pages, identifying nine domain types, and revealing the dominance of illegal marketplaces and domain siloing behaviors.

Contribution

It offers the first broad, systematic evaluation of Tor content using extensive crawling, topic, and network analysis, expanding beyond prior studies focused on illicit content.

Findings

Nine domain types identified, including illegal marketplaces.

Some domain types intentionally silo themselves from others.

Illegal drug and service marketplaces are the dominant Tor domains.

Abstract

Tor is among most well-known dark net in the world. It has noble uses, including as a platform for free speech and information dissemination under the guise of true anonymity, but may be culturally better known as a conduit for criminal activity and as a platform to market illicit goods and data. Past studies on the content of Tor support this notion, but were carried out by targeting popular domains likely to contain illicit content. A survey of past studies may thus not yield a complete evaluation of the content and use of Tor. This work addresses this gap by presenting a broad evaluation of the content of the English Tor ecosystem. We perform a comprehensive crawl of the Tor dark web and, through topic and network analysis, characterize the types of information and services hosted across a broad swath of Tor domains and their hyperlink relational structure. We recover nine domain types defined by the information or service they host and, among other findings, unveil how some types of domains intentionally silo themselves from the rest of Tor. We also present measurements that (regrettably) suggest how marketplaces of illegal drugs and services do emerge as the dominant type of Tor domain. Our study is the product of crawling over 1 million pages from 20,000 Tor seed addresses, yielding a collection of over 150,000 Tor pages. We make a dataset of the intend to make the domain structure publicly available as a dataset at https://github.com/wsu-wacs/TorEnglishContent.

Tables

Table 1. Table 1. List of 10 most probable words per topic and their label

Label	top-10 words
Directory	Search, Database, Address, Tor, Browse, URL, Directories, Link, Site, Dir
Bitcoin	Btc, Bitcoin, Blockchain, Transaction, Wallet, Deposit, Coin, Buy, Anonymity, Hidden
News	Information, Newspaper, News, Tor, Events, Censorship, Web, Press, Tor, Comment
Email	PM, Privacy, Massage, Mail, GPG, Cypherpunk, AES, Webmail, IRC, Darkmail
Multimedia	Copyright, Book, Video, Music, Free, pdf, Library, TV, FLAC, Paper
Shopping	Supplier, Market, Hidden, Commerce, Product, buy, Order, Price, Marketplace, Money
Forum	NSFW, Invite, Friend, Group, Share, Private, BBS, IM, Chat, Forum
Gambling	Online, Gambling, game, Casino, Lottery, Roulette, Table, Value, Money, Play
Dream market	Marketplace, Online, Buy, Register, Dream, Support, Hidden, Account, User, Tor

Table 2. Table 2. Summary statistics of the domain network

Statistic	Value
Domain count ($\left|V\right|$)	1,766
Hyperlink count ($\left|E\right|$)	5,523
Mean degree ($\overline{k}$)	12
Max degree ($\max k$)	389
Density ($\rho$)	0.0064
W.C.C. Count ($\left|C^w\right|$)	25
S.C.C. Count ($\left|C^s\right|$)	756
Max W.C.C. Size ($\max C_i^w$)	955 (54%)
Max S.C.C. Size ($\max C_i^s$)	13 (0.73%)

Table 3. Table 4. Modularity score of each topic community

Community	$M$
Dream market	0.452
Directory	0.068
Forum	0.034
Email	0.032
Gambling	0.024
News	0.019
Multimedia	0.017
Shopping	0.012
Bitcoin	0.002

Table 4. Table 5. Power-tail distribution hypothesis test results

Community	In-degree distribution	Out-degree distribution
Bitcoin	$p=.7623,\alpha =2.69$	$p=.0114$
Forum	$p=.8996,\alpha =3.01$	$p=.0001$
Email	$p=.3377,\alpha =2.08$	$p=.0086$
News	$p=.0344$	$p=.5681,\alpha =2.73$
Directory	$p=.3407,\alpha =2.65$	$p=.0002$
Shopping	$p=.2021,\alpha =2.85$	$p=.0001$
Gambling	$p=.0002$	$p=.0002$
Multimedia	$p=.0003$	$p=.0002$
Dream Market	$p=.0005$	$p=.0007$

Equations

$$C(t;W^{(t)})=\sum _{i=2}^n\sum _{j=1}^{i-1}log\frac{F_d(w_i^{(t)},w_j^{(t)})+1}{F_d(w_j^{(t)})}$$

$$\arg \underset{T}{\min }\left|\sum _{t=1}^T\frac{C(t;W^{(t)})}{T}\right|$$

$$\gamma (\zeta ,t)=\{\begin{array}{ll}\frac{\sum _{v_x,v_y\in W^{(t)},x<y}{I}_i^{xy}}{\left(\genfrac{}{}{0ex}{}{n}{2}\right)},& \text{if }\zeta \notin W^{(t)}\\ \frac{\sum _{v_x,v_y\in W^{(t)},x<y}{I}_i^{xy}}{\frac{(n-1)(n-2)}{2}},& \text{otherwise}\end{array}$$

$$I_i^{(xy)}=\frac{1}{2}\sum _{v_j\in V_t}{A}_{ij}|T_{ij}-T_{iy}-T_{jx}+T_{jy}|$$

$$R_{\kappa }=\frac{S_1}{S_2}=\frac{{\sum }_{i=0}^{|V|}i\cdot C_i^{(\kappa )}}{{\sum }_{i=0}^{|V|}i|V|-{\sum }_{i=0}^{|V|}{i}^2}=\frac{6{\sum }_{i=0}^{|V|}i\cdot C_i^{(\kappa )}}{|V|(|V|+1)(|V|-1)}$$

$$M=\frac{1}{2|E|}\sum _{i,j}\left(A_{ij}-\frac{d_i{d}_j}{2|E|}\right){\Delta }_{typ{e}_i=typ{e}_j}$$

Full text

Tor, Content Analysis, Structural Analysis

††copyright: none††copyright: none

1. Introduction

The deep web defines content on the World Wide Web that cannot or has not yet been indexed by search engines. Services of great interest to parties that want to be anonymous online is a subset of the deep web called dark nets: networks running on the Internet that require unique application layer protocols and authorization schemes to access. While many dark nets exist (i.e. I2P (Chertoff and Simon, 2015), Riffle (Henri, 2017), and Freenet (Clarke et al., 2010)), Tor (Syverson et al., 2004) has emerged as the most popular (Zulkarnine et al., 2016; Sanchez-Rola et al., 2017). It is used as a tool for circumventing government censorship (Loesing et al., 2010), for releasing information to the public (Jardine, 2018), for sensitive communication between parties (Zhang et al., 2010; Weimann, 2016), and as an (allegedly) private space to buy and sell goods and services (Maddox et al., 2016; Rhumorbarbe et al., 2018).

It is an open question whether the fundamental and often necessary protections that Tor provides its users is worth its cost: the same features that protect the privacy of virtuous users also make Tor an effective means to carry out illegal activities and to evade law enforcement. Various positions on this question have been documented (Hayes et al., 2018; Pannu et al., 2018; Kassab and Rosen, 2019), but empirical evidence is limited to studies that have crawled, extracted, and analyzed specific subsets of Tor based on the type of hosted information, such as drug trafficking (Dolliver et al., 2018), homemade explosives (Iliou et al., 2016), terrorist activities (Chen et al., 2008), or forums (Zhang et al., 2010).

A holistic understanding of how Tor is utilized and its structure as an information ecosystem cannot be gleamed by surveying this body of past work. This is because previous studies focus on a particular subset of this dark web and take measurements that aim to answer unique collections of hypotheses. But such a holistic understanding of Tor’s utilization and ecosystem is crucial to answer broader questions about this dark web, such as: How diverse is the information and the services provided on Tor? Is the argument that the use of Tor to buy and sell illicit goods and services and to enable criminal activities valid? Is the domain structure of Tor ‘siloed’, in the sense that the Tor domain hyperlink network is highly modular conditioned on content? Does the hyperlink network between domains exhibit scale-free properties? Answers to such questions can yield an understanding of the kinds of services and information available on Tor, reveal the most popular and important (from a structural perspective) services it provides, and enable a comparison of Tor against the surface web and other co-reference complex systems.

Towards this end, we present a quantitative characterization of the types of information available across a large swath of English language Tor webpages. We conduct a massive crawl of Tor starting from 20,000 different seed addresses and harvest only the html page of each visited address111For privacy purposes no embedded resources of any kind, including images, scripts, videos, or other multimedia, were downloaded in our crawl.. Our crawl encompasses over 1 million addresses, of which 150,473 are hosted on Tor and the remaining 1,085,960 returns to the visible web. We focus on 40,439 Tor pages belonging to 3,347 English language domains and augment LDA with a topic-labeling algorithm that uses DBpedia to assign semantically meaningful labels to the content of crawled pages. We further extract and study a logical network of English Tor domains connected by hyperlinks. We limit our study to the subset of Tor domains with information written in English in an attempt to control for any variability in measurements and insights that could be caused by the confluence of information posted in unique languages, structure, and possibly unique subdomains of Tor (for example, VPN services specifically targeting users in countries with government controlled censorship or marketplaces exclusively for users living in a particular country). We summarize our insights to the following research questions:

• •

RQ1: Is the information and services of Tor diverse?

• –

We find that Tor services can be described by just nine types. Over 50% of all domains discovered are either directories to other Tor domains, or serve as marketplaces to buy and sell goods and services. Just 24% of all Tor domains are used to publicly post, privately send, or to discover information anonymously. We do find, however, that different types of domains relate to each other in unique ways, including marketplaces that enable payment by forcing gameplay on a gambling site, and that domains involving money transactions have a surprisingly weak reliance to Tor Bitcoin domains.

• •

RQ2: What are the ‘core’ services of Tor? Is there even a core?

• –

An importance analysis from some centrality metrics finds the Dream market to be the most structurally important, “core” service Tor provides by a wide margin. The Dream market is the largest marketplace for illicit goods and services on Tor. Directory sites to find and access Tor domains have dominant betweenness centrality, making them important sources for Tor browsing.

• •

RQ3: How siloed are Tor information sources and services?

• –

A connectivity analysis shows how Tor services tend to isolate themselves which makes them difficult to discover by simple browsing and implies the need for a comprehensive seed list of domains for data collection on Tor. We also find patterns suggesting competitive and cooperative behavior between domains depending on their domain type, and that Tor is not particularly introspective: few news domains reference a large number of other domains across this dark web.

• •

RQ4: How can the structure of Tor domain hyperlinks be modeled? What are the implications of Tor’s domain connectivity?

• –

We find that the hyperlink network of Tor domains does not follow the scale-free structure observed in other sociotechnological systems or the surface web. Also, investigating within content communities denotes a power-tailed pattern in the in-degree distributions of about half of all Tor domains.

To the best of our knowledge, this study reports on the largest measurements of Tor taken to date, and is the first to study the relationship between Tor domains conditioned on the type of information they hold. We publish the hyperlink structure of this massive dataset to the public at: https://github.com/wsu-wacs/TorEnglishContent. Contrasting our findings for subsets of Tor written in other languages will be an exciting direction for future work.

This paper is organized as follows: Section 2 discusses related work on characterizing and evaluating Tor. Section 3 presents the procedure used for data collection and processing. Section 4 provides an evaluation on Tor content while Section 5 describes the logical network of the domains connected by hyperlinks and presents the analyses results. Finally, Section 6 summarizes the main conclusions and discusses the future work.

2. Related Work

Previous research on Tor have focused on characterizing particular types of hosted content, on traffic-level measurements, and on understanding the security, privacy, and topological properties of Tor relays at the network layer. Towards understanding types of content on Tor, Dolliver et al. use geovisualizations and exploratory spatial data analyses to analyze distributions of drugs and substances advertised on the Agora Tor marketplace (Dolliver et al., 2018). Chen et al. seek an understanding of terrorist activities by a method incorporating information collection, analysis, and visualization techniques from 39 Jihad Tor sites (Chen et al., 2008). Mörch et al. analyze 30 Tor domains to investigate the accessibility of information related to suicide (Mörch et al., 2018). Dolliver et al. crawls Silk Road 2 with the goal of comparing its nature in drug trafficking operations with that of the original site (Dolliver, 2015). Dolliver et al. also conduct an investigation on psychoactive substances sold on Agora and the countries supporting this dark trade (Dolliver and Kuhns, 2016). Other related works propose tools to support the collection of specific information, such as a focused crawler by Iliou et al. (Iliou et al., 2016), new crawling frameworks for Tor by Zhang et al. (Zhang et al., 2010), and advanced crawling and indexing systems like LIGHTS by Ghosh et al. (Ghosh et al., 2017).

Tor traffic monitoring is another related area of work. This monitoring is often done to detect security risks and information leakage on Tor that can compromise the anonymity of its users and the paths packets take. Mohaisen et al. study the possibility of observing Tor requests at global DNS infrastructure that could threaten the private location of servers hosting Tor sites, such as the name and onion address of Tor domains (Mohaisen and Ren, 2017). McCoy et al. study the clients using and routers that are a part of Tor by collecting data from exit routers (McCoy et al., 2008). Biryukov et al. analyze the traffic of Tor hidden services to evaluate their vulnerability against deanonymizing and take down attacks. They demonstrate how to find the popularity of a hidden service, harvest their descriptors in a short time, and find their guard relays. They further propose a large-scale attack to disclose the IP address of a Tor hidden service (Biryukov et al., 2013). We consider such work, operating at the traffic level, as studying Tor payloads that pass through networks, rather than in understanding the content of these payloads or of the inter-connected structure of Tor domains.

The topological properties of Tor, at physical and logical levels, are only beginning to be studied. Xu et al. quantitatively evaluated the structure of four terrorist and criminal related networks, one of which is from Tor (Xu and Chen, 2008). They find such networks are efficient in communication and information flow, but are vulnerable to disruption by removing weak ties that connect large connected components. Sanchez-Rola et al. conducted a broader structural analysis over 7,257 Tor domains (Sanchez-Rola et al., 2017). Their experiments indicate that domains are logically organized in a sparse network, and finds a surprising relation between Tor and the surface web: there are more links from Tor domains to the surface web than to other Tor domains. They also find evidence to suggest a surprising amount of user tracking performed on Tor. Part of this work extends their study by examining the structure between Tor domains conditioned on the type of information they host.

3. Dataset collection and Processing

We performed a comprehensive crawl of the Tor network to extract data for this study. Figure 1 illustrates this data collection process. We developed a multi-threaded crawler that collects the html of any Tor website reachable by a depth first search (up to depth 4) starting from a seed list of 20,000 Tor addresses. This seed list was constructed by concatenating the list used in a recent study (Sanchez-Rola et al., 2017) along those identified by the author’s manual search of $Reddit$, $Quora$, and $Ahima$, and other major surface web directories in the days predating the crawl. Although a manual list of seeds runs the inevitable risk of a crawl that can miss portions of Tor, the hidden nature of Tor websites make it unlikely for there to ever be a single authoritative directory of Tor domains. We are confident that our seed list leads to a comprehensive crawl of Tor because:  (i) The Reddit, Quora, Ahima, and surface web directories are well-known for providing up to date links to Tor domains and are often used by Tor users to begin their own searches for information. This suggest that these entry points into Tor are at worst practically useful, and at best are ideal starting points to search and find Tor domains associated with the most common uses of the service; (ii) The list adapted from (Sanchez-Rola et al., 2017) are noted to be sources commonly utilized to discover current Tor addresses. Furthermore, we assign our crawlers to cover all hyperlinks up to depth 4 from every seed page to make our data collection as comprehensive as possible. Out seed list is published online.222https://github.com/wsu-wacs/TorEnglishContent/blob/master/seeds

Because of the rapidly changing content and structure of Tor (Sanchez-Rola et al., 2017), including temporary downtime for some domains, we executed two crawls 30 days apart from each other in June and July 2018. To try to control for some variability in the up and down time of domains, the union of the Tor sites captured during the two crawls were stored for subsequent analysis. It is worth noting that we only request html and follow hyperlinks, and do not download the full content of a web page. This prevents any access control polices, request rate limiters, and crawler blockers from interrupting our data collection. A total of 1,236,433 distinct pages were captured across both crawls. The collected data was post-processed to identify English pages and to classify the crawled pages as being from the surface or from Tor. Any webpage with suffix .onion was classified as a Tor page, while the remainder are considered to be from the surface web. A language identification method proposed by (Feinerer et al., 2013) was used to remove non-English onion pages based on their text content regardless of the value set in their HTML language tag. 40,439 English Tor pages remained after this filtering. We chose to only focus on English pages to facilitate our content analysis; an evaluation of non-English pages will be the topic of future work.

3.1. Tor content discovery and labeling

We subjected the corpus of English Tor pages through an unsupervised content discovery and labeling procedure. The process runs the content of every Tor page (where content is defined as any string outside of a markdown tag) through the Latent Dirichlet Allocation (LDA) (Blei et al., 2003) and Graph-based Topic Labeling (GbTL) (Hulpus et al., 2013) algorithms to derive a collection of semantic labels representing broad topics of content on Tor. Each Tor domain is then assigned a label by the dominant topic present across the set of all webpages crawled in the domain.

3.1.1. Topic Modeling

Topic modeling (Wallach, 2006) is a method to uncover topics as hidden structures within a collection of documents. By defining a topic as a group of words occurring often together, topic modeling creates semantic links among words within the same context, and differentiates words by their meaning. LDA  (Blei et al., 2003) is a widely used unsupervised learning technique for this purpose. It models a topic $t_{j}$ ($1\leq j\leq T$) by a probability distribution $p(w_{i}|t_{j})$ over words taken from a corpus $D=\{d_{1},d_{2},\cdots,d_{N}\}$ of documents where words are drawn from a vocabulary $W=\{w_{1},w_{2},\cdots,w_{M}\}$. The probability of observing word $w_{i}$ in document $d$ is defined as $p(w_{i}|d)=\sum_{j=1}^{T}p(w_{i}|t_{j})p(t_{j}|d)$. Gibbs sampling is used to estimate the word-topic distribution $p(w_{i}|t_{j})$ and the topic-document distribution $p(t_{j}|d)$ from data.

An important hyperparameter is the number of topics $T$ that should be modeled. We set $T$ by considering the coherence (Mimno et al., 2011) of a set of topics inferred for some $T$, choosing the $T$ with the best coherence $C$. $C$ is a function of the $n$ words of each $t_{i}$ having highest probability $P(w_{j}|t_{i})$. Let $W^{(t)}=\{w_{1},\cdots,w_{n}\}$ be the set of top-$n$ most probable words from $P(w|t)$. Then $C$ is given by:

[TABLE]

where $w_{i}^{(t)},w_{j}^{(t)}\in W^{(t)}$, $F_{d}(w)$ indicates the number of documents where $w$ emerges, and $F_{d}(w_{i},w_{j})$ gives the number of documents in which both words $w_{i}^{(t)}$ and $w_{j}^{(t)}$ exist. Values closer to zero indicate higher coherence for the corresponding topic. $T$ is then chosen as the one that yields the smallest average $C$ over all topics:

[TABLE]

where the summation is ran over a model fitted to $T$ topics.

A final parameter is the minimum length of a document (e.g. Tor page) for it to be considered in the coherence calculation. We determine this length empirically by inspection of Figure 2, which gives topic coherence scores for different values of $T$ and different minimum lengths of the documents where $n=10$ is considered for each topic. The trend for $T=9$ yields the closest value to zero for document minimum length of 50 words.

3.1.2. Graph-based Topic Labeling with DBpedia

Word-topic distributions from a fitted topic model are indicative of semantically related words appearing in common contexts. A human may subsequently assign a label to each topic by manually evaluating this distribution However, the manual approach yields a subjective, possibly biased interpretation of the topics present in a corpus. Instead, we incorporate the unsupervised knowledge graph-based labeling algorithm GbTL (Hulpus et al., 2013). It utilizes the DBpedia (Bizer et al., 2009) knowledge graph (KG) which codifies Wikipedia articles and their relationships in the form of an ontology. GbTL finds a concept $\zeta$ from DBpedia that would serve as a suitable label to represent $t$.

To determine $\zeta$, GbTL defines a suitability measure $\gamma$. Before defining $\gamma$, we note that optimizing any measure over all DBpedia concepts is infeasible due to the massive size of the DBpedia ontology. Instead, GbTL considers a candidate set of possible labels for a topic $t$. The candidate set for topic $t$ are all vertices in the subgraph $G_{t}=(V_{t},E_{t})$ of DBpedia where $V_{t}$ is the set of concepts with labels identical to any word in $W^{(t)}$ along with their directed 1st and 2nd degree neighborhood and $E_{t}$ is the relations among all concepts in $V_{t}$. The choice of 2nd degree neighbors is based on (Hulpus et al., 2013) that found this setting to produce a sufficiently large candidate set of labels without adding unrelated ones. Since topic labels can be considered as assigning classes/categories to the topics, we restrict the subgraph relations to those of type rdfs:type, dcterms:subject, skos:broader, skos:broaderOf, and rdfs:sub-ClassOf. The suitability $\gamma$ of each concept $\zeta$ in $V_{t}$ is measured by its Focused Random Walk Betweenness Centrality measure (Hulpus et al., 2013). This centrality measures the average amount of time it takes for a random walker to arrive at some node starting from any other node in a network. It is computed as follows:

1. (1)

Let $L=D-A$ be the Laplacian matrix of $G_{t}$ where $A$ is the adjacency matrix of $G_{t}$ and $D$ is its diagonal degree matrix. 2. (2)

Arbitrarily remove a row and its corresponding column from $L$ and then invert it. Define $T$ as this inverse with a row and column vector of zeroes inserted at the same index the row and column was removed from $L$. 3. (3)

Define $\gamma(\zeta,t)$ as follows:

[TABLE]

where $I_{i}^{xy}$ is given as:

[TABLE]

where $i$ is the index of $A$ corresponding to $\zeta$ and $T_{ab}$ is the value at row $a$ and column $b$ of $T$. Since DBpedia is a semantic graph, we assume that vertices playing an important structural role in the graph should be representative of a concept that binds together the concepts in $W^{(t)}$. We use $\arg\max_{\zeta}\gamma(\zeta,t)$ as the label of topic $t$.

We built the topic models and derived their subsequent labels using the 7,782 English Tor web pages having more than 50 words. These 7,782 pages come from 1,766 unique onion domains and represents the data we consider in the remainder of our study.333In the remainder of the paper, for sake of brevity we will sometimes refer to the set of English Tor domains as simply ‘Tor’ or ‘Tor domains’.

4. Content Evaluation

We assigned each domain’s label by the dominant topic in a concatenation of all of its constituent pages. Application of LDA to this set of documents yielded a set of 9 topics. GbTL labeled these topics Forum, Shopping, Bitcoin, Dream market, Directory, Multimedia, News, Email service, and Gambling. Table 1 lists the 10 most probable words per topic. We supplement this with a manual evaluation of the html code of pages in each domain to better understand the meaning of each topic. For domains where we noticed a number of sub-types identified (e.g. a shopping domain that specialized in particular types of goods or services), we maintained a count of sub-type frequencies and show them in Figure 3. We elaborate on our evaluation of each type of domain below:

Shopping: (359 domains [20.32%]) Shopping domains allow visitors to purchase goods and services, including drugs, medicine, as well as consultancy and investment services. We found the dominant shopping service is to provide money transfer to and from credit cards to buy Visa cards, gift cards, and Bitcoin. Drugs, pornography, hosting Tor services, and forgery services are other types of popular shopping domains. The rest includes pages selling a variety of goods like phone, laptop, movies, and even gold.

Bitcoin: (83 domains [4.7%]) Bitcoin domains provide services for Bitcoin transactions and fund transfers to wallets. Tor Bitcoin services differ in terms of the registration process and personal information needed, registration and transaction fees, the security platforms they use, and fees compared to surface web counterparts.

Dream market: (235 domains [13.3%]) One shopping domain so large that it merited its own topic category is the Dream market. Its pages suggest a wide range of content available for sale, most of which are illicit. This includes drugs, stolen data, and counterfeit consumer goods. Many Dream market pages collected includes login and registration forms (hence words like “Register”, “Account”, and “User” emerge in the list of top-10 terms in Table 1) limiting further access for our crawlers.

Directory: (445 domains [25.19%]) Addresses in Tor are made up of meaningless combinations of digits and characters that sometimes change and are hard to memorize. Moreover, most domains may disappear after a short amount of time or move to new addresses (Sanchez-Rola et al., 2017). It is thus no surprise to find directory domains emerge in our study. Directories are a convenient way of finding Tor domains without knowing their addresses. Domains labeled as a directory include unnamed pages with lists of .onion domains along with the better known TorDir and The Hidden Wiki services, as well as search engines like DeepSearch and Ahmia.

Multimedia: (46 domains [2.6%]) Multimedia domains are sites to download and purchase multimedia products like e-books, movies, musics, games, and academic and press articles even if they are copyright protected. Among all multimedia domains, we measured 28% of them to exclusively offer articles and e-books while 22% offer free download of music or audio files, and to even obtain login information for stolen TV accounts. The remaining provides resources like hacked video game accounts, cracked software, and a mixture of the above.

Forum: (111 domains [6.3%]) Forum domains host bulletin board and social network services for Tor users to discuss ideas and thoughts. Among all forums, 72% have a range of topic discussions including information on Tor and its services, hacking, sexual health, dark net markets, weapons, drugs, and trusted Tor domains. Some forums require payment via Bitcoin to register.

News: (183 domains [10.36%]) Whereas forums facilitate interaction among Tor users, news domains host pages akin to personal weblogs where an author writes an essay and visitors can post follow-up comments. Links to Tor e-mail services are included to contact post authors. Information on current Tor services and directories is presented by most news domains, along with politics, Bitcoin, drug, and Tor security related posts.

Email service: (124 domains [7.02%]) Email domains offer communication services like email, chat room, and Tor VPNs. Email services vary in the encryption protocol they use, the advertisements they serve for other services, and policies for keeping user log files. Our investigation finds that many email domains use secure protocols like SSL, AES, and PGP to secure user accounts and messages. IRC-based chat rooms are also common. Some email services charge recurring subscription fees.

Gambling: (180 domains [10.19%]) Gambling domains offer services to bet money on games, to purchase gambling advice and consulting, and to read gambling-related news. Gambling domains have a number of links to payment processing options including Ethereum, Monero, DASH, Vertcoin, Visa, and MasterCard.

Figure 4 gives the distribution of topics assigned to each Tor domain. It illustrates how directory and shopping domains, and the Dream market dominate English domains on Tor by accounting for 58.83% of all domain types. This suggests that Tor’s main utility for users may be to browse information and shop on marketplaces that require secrecy. In contrast, domains related to the free exchange of ideas and information (a powerful and positive use-case of Tor, particular to users in countries facing Internet censorship (Jardine, 2018)), represented by Forum, Email, and News sites, account for just 23.66% of all domains. It should be noted, however that services used in countries that most benefit from the freedom of expression of Tor may not be well-represented in our sample. Gambling domains represent a surprisingly large (10.19%) percentage of domains, suggesting that people may now be turning to Tor to play online gambling games that are otherwise illegal to host in many countries around the world. Finally, and perhaps surprisingly, Bitcoin and multimedia domains respectively constitute the smallest proportion of English Tor domains. Our initial expectation was that Bitcoin domains would be popular given the prevalence of the Dream market and shopping domains where purchases are made via cryptocurrency. We postulate that sites having a Bitcoin domain may host wallets, search a blockchain, or be markets that covert currency to BTC. Such sites hence may only need to be visited infrequently. Moreover, the mainstream popularity of Bitcoin has led to reputable surface web domains offering similar Bitcoin services.

5. Domain Relationships

We next examine the structure of relationships between different sources of information on Tor. Such structural analyses realize the inter and intra-connectivity of domains on Tor conditioned by the type of information or content they host. The structural analysis also seeks to identify the topological properties of the Tor domain network to evaluate similarities and differences between the structure and formation process of Tor domains compared to the surface web and other sociotechnological systems. We build a graph where vertices are domains and a directed relation means a page in a domain has a hyperlink to a page in a different domain.

We measure simple statistics on the connectivity and connected components of the graph in Table 2 to make sense of the structure visualized in Figure 5 (note that all domains with degree [math] are excluded from the figure). The network is sparse, with only 5,523 undirected edges among 1,766 domains and a network density $\rho=0.006$. The network also has $|C^{w}|=25$ weakly connected components (W.C.C.) with the largest one having 955 domains. That only 54% of all vertices fall in the largest W.C.C. is somewhat surprising as many sociotechnological systems including the surface Web hyperlink graph exhibit a single massive connected component that the vast majority of all vertices participate in (Broder et al., 2000). This suggests that the underlying process for linking between websites is fundamentally different than in most sociotechnological systems: rather than encouraging connections to make information dissemination easier, the modus operandi of Tor domain owners may be to discourage linking to other domains, so that information on this “hidden” web becomes difficult to arbitrarily discover and disseminate. This hypothesis is further supported by examining the set of strongly connected components $C^{s}$. We measure $|C^{s}|=756$ and $\max{C^{s}_{i}}=13$, suggesting that an extraordinarily small number of domains collectively co-link with each other. It may be the case that the relatively larger $|C^{w}|$ and $\max{C^{w}_{i}}$ are simply caused by directory websites that offer links to a variety of other domains. These interesting deviations from other types of Web graphs collectively suggest that information on Tor may be intentionally isolated and difficult to discover by simple browsing. The small maximum connected component size speaks to the need of a comprehensive seed list of domains to for any comprehensive crawl of Tor.

5.1. Connectivity analysis

We next examine hyperlink relationships across domains, within domains, and the tendency of domains to link to like domains.

5.1.1. Inter-connectivity

To investigate the relationship between specific domains, we redraw the network with a carpet layout where vertices are spatially grouped in a grid by their domain type in Figure 6. We also list the sum of in- and out-degree from each community to every other community in Figure 7. There are some notable domain relationships observed in the two figures. For example, we find a small number of outgoing edges from shopping to gambling domains (Panel 1 of Figures 6 and 7). Our manual investigation of these relations find that some shopping domains actually provide customers with a method of payment by gambling, where a customer and seller play an online game to determine an amount of payment. We further note that shopping websites are isolated from the Dream market, perhaps to maintain some distinction between their offerings and the largest marketplace on Tor (Lane et al., 2018). Another interesting observation is that there are a few number of edges from shopping to Bitcoin domains. Our manual investigation indicates that in addition to some shopping domains providing in-person cash payment (usually local drug vendors), marketplaces that use cryptocurrency for payment tend to link to providers on the surface web, sometimes with instructions for the user to purchase cryptocurrency from a trading house or market, thus explaining the small number of out-going edges from marketplaces to Bitcoin domains. Such links to the surface web, however have been noted as a major Achilles heel to privacy on Tor as a cause of Tor information leakage (Sanchez-Rola et al., 2017). Links incoming to Bitcoin domains (Panel 2 of Figures 6 and 7) are dominated by directory ($\approx 53\%$) and email ($\approx 32\%$) domains. This finding defeats our intuition that marketplaces, the Dream market, and gambling sites would have been services most reliant on Bitcoin domains. We also investigated the email domains having edges to Bitcoins and found that many email services on Tor charge customers for anonymous messaging services or suggest a donation be made via Tor Bitcoin services. The sparse relation between multimedia and Bitcoin domains are due to the fact that some multimedia domains charge users for the services they provide, and of those, many utilize surface web cryptocurrency providers.

Other insights can be gleamed from the carpet and inter-domain degree distributions. For example, the outgoing connections from Dream market websites (Panel 6 of Figures 6 and 7) show that this marketplace has very few incoming or outgoing edges to other domains. Incoming links tend to originate from directory, email, and forum communities, which could be a byproduct of forum threads and email links to this secret marketplace. The Dream market thus appears to be especially siloed on the dark web, despite the fact that it represents $13.3\%$ of all domains. The outgoing edges from directory domains (Panel 7 of Figures 6 and 7) indicate that a large collection of directories on Tor may be sufficient to include sites across all major Tor domains. Email domains exhibit a similar phenomena (Panel 9 of Figures 6 and 7) where they tend to connect to all other types of domains. This suggests that Tor e-mail services are not necessarily exclusive to only marketplace or forum users, but may be a useful service for most Tor visitors. Finally, we see that news and forums have no hyperlinks to each other (Panel 4 of Figures 6 and 7 and Panel 8 of Figures 6 and 7) which implies that although both services are information providers on Tor, they work independently of each other.

5.1.2. Intra-connectivity

Next, we investigate the connectivity within each domain to evaluate how tightly knit and accessible particular types of Tor domains are among each other. This can reveal how often domains encourage their visitors to visit other domains having similar types of content. Figure 8 separately visualizes all intra-domain connections for each domain type. Perhaps unsurprisingly, Dream market domains are tightly connected compared to others, splitting into only four connected components and the smallest number of isolated domains. Shopping domains are almost entirely disconnected from each other, indicating that marketplaces may intentionally disassociate themselves with others. Gambling, multimedia, and Bitcoin domains are similarly disconnected, likely reflecting a competition for users within domains that tend to charge service fees. On the other hand, email, forum, and directory domains all exhibit a single large connected component. This implies that in email and directory communities, domains have more support from each other and refer their visitors to other similar service providers. News domains have a larger percentage of isolated domains which implies that most of the news websites work independently from each other.

We also quantify the intra-connectivity of topic domains by a Robustness coefficient $R_{\kappa}$ proposed by (Piraveenan et al., 2012). This coefficient reflects the degree to which a network shatters into multiple connected components as vertices and their incident edges are removed. To compute $R_{\kappa}$, an ordering $\mathcal{O}_{\kappa}$ is induced on the vertices of the network by a centrality measure $\kappa$ such that $\mathcal{O}_{\kappa}(k)$ gives the vertex with the $k^{th}$ highest $\kappa$ centrality. Letting $C_{i}^{(\kappa)}$ be the size of the largest connected component of the network after removing $\{\mathcal{O}_{\kappa}(1),\mathcal{O}_{\kappa}(2),...,\mathcal{O}_{\kappa}(i)\}$ and their incident edges, $R_{\kappa}$ is given by:

[TABLE]

$R_{\kappa}$ ranges in $[0,1]$ and smaller values suggest that the network shatters faster as vertices having high $\kappa$ betweenness are removed. The idea behind $R_{\kappa}$’s formulation is to quantify the change in the largest network component size $C_{i}^{(\kappa)}$ as nodes are removed from the network. The network is ‘maximally robust’ if a plot of $C_{i}^{(\kappa)}$ vs the number of nodes removed shows a simple linear decreasing trend. Then the ratio of area under such a plot for a given network, $S_{1}$, over the area under the plot for an ideal network, $S_{2}$, defines the robustness coefficient for that network. Technical details about the measure are discussed in (Piraveenan et al., 2012). Table 3 lists $R_{\kappa}$ for each intra-domain network under Betweenness $b$, closeness $c$, and degree $d$ centrality. $d$ is defined as the undirected degree of a node, $b$ is defined as the number of shortest paths in the network that a node participates in, and $c$ is defined as the inverse of the average path length from the node to all others in its connected component (Zafarani et al., 2014).

Studying Table 3 shows a correlation between $R_{\kappa}$ and the competitive or cooperative intra-domain behaviors noted above. There are some networks whose $R_{d}$ is close to $R_{c}$, namely those communities that are sparse networks (shopping, Bitcoin, news, gambling, multimedia), where competition for users and attention is natural. For domains where $R_{c}$ and $R_{d}$ are differentiated (directory, email, forum, Dream market), their structure has fewer connected components. This is best illustrated by Dream market where the difference of these coefficients has its maximum value, while exhibiting the fewest connected components. Also, we see that for networks with more supportive interlinking among domains (directory, email, and forum) the difference between $R_{c}$ and $R_{d}$ are larger.

In comparing $R_{b}$ between domains, we observe two groups with similarly small (shopping, Bitcoin, news, Dream market, gambling) and large (multimedia, directory, forum, email) values. This suggests that the intra-connectivity of domains in the first group is dependent on a small percentage of domains, or has a high number of isolated domains which have zero Betweenness centrality. On the other hand, high $R_{b}$ in the second category implies that their intra-connectivity is more robust to domain failures.

It is worth mentioning that for Dream market, $R_{b}$ and $R_{c}$ are significantly different due to the separated connected components also seen in Figure 5. Since a high percentage of domains in this community are located in a single connected component, their closeness centralities will be greater than zero. On the other hand, the separated connected components can cause low values for Betweenness centrality since this metric is based on paths between pairs of domains. This finding also indicates that based on the information we have from home pages of Dream markets, there exist separated Dream market communities that link to similar types of goods and services.

5.1.3. Modularity

Finally, we study the modularity of the network as a means to understand the relationship between the inter- and intra-domain connectivity conditioned on the content type of the domain. A domain will have high modularity if it tends to link to pages of the same content type, and low modularity if it tends to link to pages of a different type. The modularity $M$ of a network is given as (Zafarani et al., 2014):

[TABLE]

where $A$ is a binary network adjacency matrix with $A_{ij}=1$ if $v_{i}$ and $v_{j}$ are adjacent and $d_{i}$ and $d_{j}$ are the undirected degree of nodes $i$ and $j$, and $\Delta_{E}$ is an indicator that returns $1$ if statement $E$ is true and [math] otherwise. Table 4 shows that Dream market domains exhibit highest modularity by a wide margin, reinforcing the idea that Dream market domains are largely siloed from all other domains in the Tor ecosystem. Directories have lower but non-negligible modularity, leaving the impression that directory domains weakly cooperate by linking to each other. The remaining domains have substantially lower modularity; thus the majority of Tor domains strongly prefer to link to other types. This suggests that Tor domains prefer to remain isolated within their community of like domains, electing not to link to domains that offer the same type of information or services.

5.2. Importance analysis

We further examine the “importance” of particular domains as defined by various measures of network centrality. The centrality analysis only considers domains having in- or out-degree $\geq 1$. We show the CDF of betweenness centralities $b_{c}$ of domains in Figure 9a. It shows how virtually every Tor domain has a betweenness centrality near or lower than $0.05$; in fact there are only $6$ domains with betweenness centrality greater than $0.05$. These domains are the directory $HiddenWiki$ ($b_{c}=0.4$) (matching previous reports suggesting this domain is the principle Tor directory (Chertoff and Simon, 2015)), the Email domain $TorBox$ ($b_{c}=0.17$) and $VFEmail$ ($b_{c}=0.06$), a Dream market domain ($b_{c}=0.07$), and two directories in the $TorWiki$ domain ($b_{c}=0.19$ and $0.11$ respectively). An attack, removal, or failure of such directories may thus directly impact the number of Tor domains reachable by a casual browser exploring this dark web. These domains may further be crucial entry points for probes or crawlers seeking to map the structure of Tor.

Figure 9b shows the distribution of eigenvector centralities across Tor domains. We use a histogram rather than a CDF plot to better illustrate the variability between centrality values. Eigenvector centrality (Zafarani et al., 2014) describes the structural importance of a node as a function of the importance of its neighboring nodes. The eigenvector centrailty of vertex $v_{i}$ is given by the $i^{th}$ component of the eigenvector of $A$ whose corresponding eigenvalue is largest. We find a heavily skewed distribution of eigenvector centralities where a majority are close to zero. Further investigation revealed that all domains with eigenvector centrality $\geq 0.2$ are part of the Dream market. Although high eigenvector centrality does not correlate with its relative popularity or frequency of visits from users, the naturally developed organization of Tor’s domain structure places the Dream market as the most meaningful Tor domain by a wide margin (note that the highest eigenvector centrality of a non-Dream market domain is only $0.04$). This establishes the Dream market as the most structurally important, “core” service Tor provides. The inlet of Figure 9b gives a sense of the distribution for the remaining domains. Here, the distribution exhibits a number of modes corresponding to connected components of the network that is disconnected from the Dream market. The especially low eigenvector centralities of these domains are further indicative of the significance of the Dream market’s structural importance in Tor.

Figure 9c shows the distribution of closeness centralities. Like eigenvector centrality, we find a division of domains by those that exhibit extremely low or high scores, but here the majority of domains have very high closeness centrality. It is interesting to find that most domains exhibit a high closeness centrality despite intra-connectivity analysis from Section 5.1.2 suggesting that the intra-connectivity of Tor domains is sparse and has a small largest W.C.C. This outcome is likely the product of the directories $HiddenWiki$ and $TorWiki$ having high betweenness centrality that enables many pairs of domains to be few hops away from each other via these directories. This underscores the central importance of directory domains to connect Tor pages across domains, and the fact that Tor domains tend to remain undiscoverable without directories.

5.3. Scale-free structure

We also investigate signs that the hyperlink structure of domains, and structure within domains, take on the same scale-free structure seen in many other sociotechnological systems (Barabási and Bonabeau, 2003) including the hyperlink structure of the surface web (Broder et al., 2000).

Figure 10 show the CCDF of the degree distributions on log-log scale. The in-degree distribution does not exhibit a straight line pattern indicative of a power-law, with a rapid drop in the CCDF occurring in the body of the distribution around an in-degree of 10. The out-degree distribution takes on a bimodal pattern, with a set of domains having degree less than 10 and another with degrees between 10 and 100. The distribution’s patterns may be explained by the variety of inter- and intra-connectivity patterns observed within each of the domains studied in Sections 5.1.1 and 5.1.2. To quantitatively confirm the distribution is not a power-law we apply Clauset et al. hypothesis test presented in (Clauset et al., 2009). The test checks the null hypothesis $H_{0}$: the network degree distribution is power-tailed against the alternative that it is not power-tailed, and provides an estimate of the power-law exponent $\alpha$ under $H_{0}$. The test leaves little doubt that the in- and out-degree distributions of the network (Figure 10) are not power-tailed with $p=0.0001$, $p=0.056$, respectively. If we consider the edges to be undirected we still have evidence to reject $H_{0}$ with $p=0.0003$. These measurements confirm the analysis presented in (Sanchez-Rola et al., 2017) that the hyperlink structure of Tor domains does not have the same scale-free structure as the surface web.

Noting the variety of intra-connectivity patterns discussed in Section 5.1.2, we also check if the degree distribution of sites within each domain are power-tailed. These measurements include intra-domain connections as well as those connections incident to a different domain. We list the $p$-value of the test for the in- and out-degree of each domain in Table 5 and include the estimate of $\alpha$ when $H_{0}$ cannot be rejected. Interestingly, we note that it is only for the in-degree distributions of Bitcoin, forum, email, shopping, and directory domains, and the out-degree of the news domain, where there is insufficient evidence to reject $H_{0}$. The popularity of about half of all Tor domains (where popularity is defined by an incoming hyperlink) thus has a power-tailed pattern suggesting that a a small number of Bitcoin, forum, email, shopping, and directories are linked to many times more frequently than is typical. That the news domain is the only one with a power law out-degree distribution may suggest the presence of a small number of highly active news sites that offer posts discussing a far wider variety of other Tor domains compared to other news domains. The majority of news domains may thus focus on a specific topic, or are otherwise used to discuss events outside of the Tor network.

6. Conclusions and Future Work

This paper presented a broad overview of the content of English language Tor domains captured in a large crawl of the Tor network. The paper makes revelations about not the physical or logical (hyperlink) structure of Tor, but of the particular domains of information or services hosted on the service and the structure between and within such domains. Such birds-eye insights, and especially those related to the inter-connectivity of domains, cannot be acquired by synthesizing the existing low-level content analysis work that focuses on a single type of Tor domain. Content analysis carried out by LDA and GbTL, using measures of topic coherence and label suitability, identified just nine principal types of domains. Manual analysis of each domain was done to describe the meaning of each topic, and any standout ‘sub-types’ seen within them. Over half of all domains constitute site directories or marketplaces to purchase and sell goods or services, with money tansfer, drugs, and pornography servicing as the most popular types of marketplaces. Our measurements identified the Dream market as perhaps the ‘core’ service of Tor, as Dream market domains exhibit especially high closeness and eignevector centralities. The inter-connectivity of the Tor domain network is surprisingly sparse with a small maximum W.C.C. but interesting domain inter-connection patterns discussed in Section 5.1.1. Patterns in the intra-connectivity structure are further indicative of levels of cooperation (where some pages hyperlink to pages in the same domain) and competition (where pages in a domain are more likely to isolate themselves from pages in the same domain) that may be measured by robustness coefficients $R_{\kappa}$ for varying centrality scores $\kappa$. We further note evidence for rejecting the hypothesis that the global domain structure is scale-free, yet there is insufficient evidence in the in-degree distributions of some domain intra-networks and the out-degree distribution of the news domain intra-network to conclude that these subnetworks are not power law. This is indicative of different underlying processes that form connections in different intra-domain networks.

Future work can expand this study further by replicating our analysis on Tor pages of different languages (chinese, russian, persian), and then by studying topic inter-connectivity between language domains, can yield a global perspective into the Tor content ecosystem. Such analysis can further be replicated on other dark web services to better understand why these lesser known services are used. Another direction of future work is to study the evolution of the topics and communities over time using a richer topic extraction analysis based on algorithms discussed in (Hofmann, 2017; Callon et al., 1983). Finally, the study can be combined with a modern crawl of similar domains on the surface web to be able to directly compare and contrast surface and dark web hyper-link structure. Such a comparative analysis could shed light around the differences in use and information between the surface and dark web.

Acknowledgements.

This paper is based on work supported by the National Science Foundation (NSF) under Grant No. 1464104. Any opinions, findings, and conclusions or recommendations expressed are those of the author(s) and do not necessarily reflect the views of the NSF.