Fingerprint empowers developers to stop online fraud at the source

Fingerprint

How to Use Fingerprint Scanners

Authors:
(1) Todd K. Moon, Electrical and Computer Engineering Department, Utah State University, Logan, Utah;
(2) Jacob H. Gunther, Electrical and Computer Engineering Department, Utah State University, Logan, Utah. Table of Links Abstract and 1 Introduction and Background 2 Statistical Parsing and Extracted Features 3 Parse Tree Features 4 Classifier 5 Dimension Reduction 6 The Federalist Papers 6.1 Sanditon 7 Conclusions, Discussion, and Future Work A. A Brief Introduction to Statistical Parsing B. Dimension Reduction: Some Mathematical Details References Abstract Over the years there has been ongoing interest in detecting authorship of a text based on statistical properties of the text, such as by using occurrence rates of noncontextual words. In previous work, these techniques have been used, for example, to determine authorship of all of The Federalist Papers. Such methods may be useful in more modern times to detect fake or AI authorship. Progress in statistical natural language parsers introduces the possibility of using grammatical structure to detect authorship. In this paper we explore a new possibility for detecting authorship using grammatical structural information extracted using a statistical natural language parser. This paper provides a proof of concept, testing author classification based on grammatical structure on a set of “proof texts,” The Federalist Papers and Sanditon which have been as test cases in previous authorship detection studies. Several features extracted from the statistical natural language parser were explored: all subtrees of some depth from any level; rooted subtrees of some depth, part of speech, and part of speech by level in the parse tree. It was found to be helpful to project the features into a lower dimensional space. Statistical experiments on these documents demonstrate that information from a statistical parser can, in fact, assist in distinguishing authors. 1 Introduction and Background There has been considerable effort over the years related to using statistical methods to identify authorship of texts, based on examples from candidate authors, in what is sometimes called “stylometry” or “author identification.” Statistical analysis of documents goes back to Augustus de Morgan in 1851 [1, p. 282], [2, p. 166], who proposed that word length statistics might be used to determine the authorship of the Pauline epistles. Stylometry was employed as early as 1901 to explore the authorship of Shakespeare [3]. Since then, it has been employed in a variety of literary studies (see, e.g. [4, 5, 6]), including twelve of The Federalist Papers which were of uncertain authorship [7] — which we re-examine here — and an unfinished novel by Jane Austen —which we also re-examine here. Information theoretic techniques have also been used more recently [8]. Earlier work in stylometry has been based on “noncontextual words,” words which do not convey the primary meaning of the text, but which act in the background of the text to provide structure and flow. Noncontextual words are at least plausible, since an author may address a variety of topics, so particular distinguishing words are not necessarily revealing of authorship. In noncontextual word studies, a set of most common words noncontextual is selected [2], and documents are represented by word counts, or ratios of word counts to document length. A review of the statistical methods is in [9]. As a variation, sets of ratios of counts of noncontextual word patterns to other word patterns are also employed [10]. Statistical analysis based on author vocabulary size vs. document length — the “vocabulary richness” — has also been explored [11]. For other related work, see [12, 13, 14, 15] A more recent paper [16] considers the effectiveness of a wide variety of feature sets. Feature sets considered there include: vectors comprising frequencies of pronouns; function words (that is, articles, pronouns, particles, expletives); part of speech (POS); most common words; syntactic features (such as noun phrase, or verb phrase); or tense (e.g. use of present or past tense); voice (active of passive). In [16], feature vectors are formed from combinations of histograms, then reduced in dimensionality using a two-stage process of principle component analysis [17] followed by dimension reduction using linear discriminant analysis (LDA). In their LDA, the within-cluster scatter matrix is singular (due to the high dimension of the feature vectors relative the number of available training vectors), so their scatter matrix is regularized. To test this, the authors consider a range of regularization parameters, selecting one which gives the best performance. More recent work [18] mentions the survey in [15] in which commonly used features in the authorship field are word and character n-grams. As noted, there are risks the statistical methods might be biased by topic-related patterns. As [18] observe, “an authorship classifier (even a seemingly good one) might end up unintentionally performing topic identification if domain-dependent features are used. ... In order to avoid this, researchers might limit their scope to features that are clearly topic-agnostic, such as function words or syntactic features.” The work presented here falls in the latter category, making use of grammatical structures statistically extracted from the text. These appear to be difficult to spoof. Examination of other recent works [19, 20] indicate that there is ongoing interest in author identification methods, but none making use of the grammatical structures use here; there is a tendency to rely more on traditional n-grams. In this work the feature vectors are obtained using tree information from parse trees from a natural language parsing tool [21]. These features were not among the features considered in [16]. The grammatical structures are, it seems, more subtle than simple counts of classes of words, and hence may be less subject to spoofing or topic bias, since it seems unlikely that an author intending to imitate another would be able to coherently track complicated patterns of usage, and the features do not include any words from the documents. It is found that the tree-based features perform better than the POS features on the test data considered. The feature vectors so obtained can be of very high dimension, so dimension reduction is also performed here. However, to deal with the singularity of the within-cluster scatter matrix, a generalized SVD approach is used, which avoids the need to select a regularization parameter. This paper provides a proof-of-concept of these tree-based features to distinguish authorship by applying them to documents which have been previously examined, The Federalist Papers and Sanditon. The ability to classify by authorship is explored for several feature vectors obtained from the parsed information. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Todd K. Moon, Electrical and Computer Engineering Department, Utah State University, Logan, Utah; (2) Jacob H. Gunther, Electrical and Computer Engineering Department, Utah State University, Logan, Utah. Authors: Authors: (1) Todd K. Moon, Electrical and Computer Engineering Department, Utah State University, Logan, Utah; (2) Jacob H. Gunther, Electrical and Computer Engineering Department, Utah State University, Logan, Utah. Table of Links Abstract and 1 Introduction and Background Abstract and 1 Introduction and Background 2 Statistical Parsing and Extracted Features 2 Statistical Parsing and Extracted Features 3 Parse Tree Features 3 Parse Tree Features 4 Classifier 4 Classifier 5 Dimension Reduction 5 Dimension Reduction 6 The Federalist Papers 6 The Federalist Papers 6.1 Sanditon 6.1 Sanditon 7 Conclusions, Discussion, and Future Work 7 Conclusions, Discussion, and Future Work A. A Brief Introduction to Statistical Parsing A. A Brief Introduction to Statistical Parsing B. Dimension Reduction: Some Mathematical Details B. Dimension Reduction: Some Mathematical Details References References Abstract Over the years there has been ongoing interest in detecting authorship of a text based on statistical properties of the text, such as by using occurrence rates of noncontextual words. In previous work, these techniques have been used, for example, to determine authorship of all of The Federalist Papers. Such methods may be useful in more modern times to detect fake or AI authorship. Progress in statistical natural language parsers introduces the possibility of using grammatical structure to detect authorship. In this paper we explore a new possibility for detecting authorship using grammatical structural information extracted using a statistical natural language parser. This paper provides a proof of concept, testing author classification based on grammatical structure on a set of “proof texts,” The Federalist Papers and Sanditon which have been as test cases in previous authorship detection studies. Several features extracted from the statistical natural language parser were explored: all subtrees of some depth from any level; rooted subtrees of some depth, part of speech, and part of speech by level in the parse tree. It was found to be helpful to project the features into a lower dimensional space. Statistical experiments on these documents demonstrate that information from a statistical parser can, in fact, assist in distinguishing authors. 1 Introduction and Background There has been considerable effort over the years related to using statistical methods to identify authorship of texts, based on examples from candidate authors, in what is sometimes called “stylometry” or “author identification.” Statistical analysis of documents goes back to Augustus de Morgan in 1851 [1, p. 282], [2, p. 166], who proposed that word length statistics might be used to determine the authorship of the Pauline epistles. Stylometry was employed as early as 1901 to explore the authorship of Shakespeare [3]. Since then, it has been employed in a variety of literary studies (see, e.g. [4, 5, 6]), including twelve of The Federalist Papers which were of uncertain authorship [7] — which we re-examine here — and an unfinished novel by Jane Austen —which we also re-examine here. Information theoretic techniques have also been used more recently [8]. Earlier work in stylometry has been based on “noncontextual words,” words which do not convey the primary meaning of the text, but which act in the background of the text to provide structure and flow. Noncontextual words are at least plausible, since an author may address a variety of topics, so particular distinguishing words are not necessarily revealing of authorship. In noncontextual word studies, a set of most common words noncontextual is selected [2], and documents are represented by word counts, or ratios of word counts to document length. A review of the statistical methods is in [9]. As a variation, sets of ratios of counts of noncontextual word patterns to other word patterns are also employed [10]. Statistical analysis based on author vocabulary size vs. document length — the “vocabulary richness” — has also been explored [11]. For other related work, see [12, 13, 14, 15] A more recent paper [16] considers the effectiveness of a wide variety of feature sets. Feature sets considered there include: vectors comprising frequencies of pronouns; function words (that is, articles, pronouns, particles, expletives); part of speech (POS); most common words; syntactic features (such as noun phrase, or verb phrase); or tense (e.g. use of present or past tense); voice (active of passive). In [16], feature vectors are formed from combinations of histograms, then reduced in dimensionality using a two-stage process of principle component analysis [17] followed by dimension reduction using linear discriminant analysis (LDA). In their LDA, the within-cluster scatter matrix is singular (due to the high dimension of the feature vectors relative the number of available training vectors), so their scatter matrix is regularized. To test this, the authors consider a range of regularization parameters, selecting one which gives the best performance. More recent work [18] mentions the survey in [15] in which commonly used features in the authorship field are word and character n-grams. As noted, there are risks the statistical methods might be biased by topic-related patterns. As [18] observe, “an authorship classifier (even a seemingly good one) might end up unintentionally performing topic identification if domain-dependent features are used. ... In order to avoid this, researchers might limit their scope to features that are clearly topic-agnostic, such as function words or syntactic features.” The work presented here falls in the latter category, making use of grammatical structures statistically extracted from the text. These appear to be difficult to spoof. Examination of other recent works [19, 20] indicate that there is ongoing interest in author identification methods, but none making use of the grammatical structures use here; there is a tendency to rely more on traditional n-grams. In this work the feature vectors are obtained using tree information from parse trees from a natural language parsing tool [21]. These features were not among the features considered in [16]. The grammatical structures are, it seems, more subtle than simple counts of classes of words, and hence may be less subject to spoofing or topic bias, since it seems unlikely that an author intending to imitate another would be able to coherently track complicated patterns of usage, and the features do not include any words from the documents. It is found that the tree-based features perform better than the POS features on the test data considered. The feature vectors so obtained can be of very high dimension, so dimension reduction is also performed here. However, to deal with the singularity of the within-cluster scatter matrix, a generalized SVD approach is used, which avoids the need to select a regularization parameter. This paper provides a proof-of-concept of these tree-based features to distinguish authorship by applying them to documents which have been previously examined, The Federalist Papers and Sanditon. The ability to classify by authorship is explored for several feature vectors obtained from the parsed information. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Your Writing Has a Fingerprint—And This Cutting Edge AI Model Can Identify It

Din skrift har et fingeraftryk - og denne banebrydende AI-model kan se det

Tu escritura tiene una huella digital y este modelo de inteligencia artificial de vanguardia puede detectarla

לכתיבה שלך יש טביעת אצבע - ודגם הבינה המלאכותית החדיש הזה יכול לזהות אותה

あなたの筆跡には指紋がある。最先端の AI モデルがそれを検出できる

Сиздин жазууңузда манжа изи бар жана бул эң алдыңкы AI модели аны байкай алат

Kokoma na yo Ezali na empreinte ya mosapi —Mpe Modèle Oyo ya AI ya sika Ekoki Komona Yango

ການຂຽນຂອງເຈົ້າມີລາຍນິ້ວມື - ແລະຕົວແບບ AI ທີ່ມີການຕັດແຂບນີ້ສາມາດເຫັນໄດ້

Tulisan Anda Mempunyai Cap Jari—Dan Model AI Cutting Edge Ini Dapat Mengesannya

Sua escrita tem uma impressão digital — e este modelo de IA de ponta pode identificá-la

การเขียนของคุณมีลายนิ้วมือ และโมเดล AI ล้ำสมัยนี้สามารถระบุลายนิ้วมือของคุณได้

Matsalwa Ya Wena Ma Ni Xikombiso Xa Tintiho—Naswona Modelo Lowu Wa AI Wa Mahungu Lamanene Wu Nga Swi Vona

Ubhalo lwakho lunophawu lweminwe-kwaye lo mzekelo wokusika we-AI unokuwubona

Authors:
(1) Diwen Xue, University of Michigan;
(2) Reethika Ramesh, University of Michigan;
(3) Arham Jain, University of Michigan;
(4) Arham Jain, Merit Network, Inc.;
(5) J. Alex Halderman, University of Michigan;
(6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad;
(7) Roya Ensaf, University of Michigan. Table of Links Abstract and 1 Introduction 2 Background & Related Work 3 Challenges in Real-world VPN Detection 4 Adversary Model and Deployment 5 Ethics, Privacy, and Responsible Disclosure 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6.2 ACK-based Fingerprinting 6.3 Active Server Fingerprinting 6.4 Constructing Filters and Probers 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7.2 Choice of Observation Window N 7.3 Effects of Packet Loss 7.4 Server Churn for Asynchronous Probing 7.5 Probe UDP and Obfuscated OpenVPN Servers 8 Real-world Deployment Setup 9 Evaluation & Findings and 9.1 Results for control VPN flows 9.2 Results for all flows 10 Discussion and Mitigations 11 Conclusion 12 Acknowledgement and References Appendix A Appendix This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Diwen Xue, University of Michigan; (2) Reethika Ramesh, University of Michigan; (3) Arham Jain, University of Michigan; (4) Arham Jain, Merit Network, Inc.; (5) J. Alex Halderman, University of Michigan; (6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad; (7) Roya Ensaf, University of Michigan. Authors: Authors: (1) Diwen Xue, University of Michigan; (2) Reethika Ramesh, University of Michigan; (3) Arham Jain, University of Michigan; (4) Arham Jain, Merit Network, Inc.; (5) J. Alex Halderman, University of Michigan; (6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad; (7) Roya Ensaf, University of Michigan. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Background & Related Work 2 Background & Related Work 3 Challenges in Real-world VPN Detection 3 Challenges in Real-world VPN Detection 4 Adversary Model and Deployment 4 Adversary Model and Deployment 5 Ethics, Privacy, and Responsible Disclosure 5 Ethics, Privacy, and Responsible Disclosure 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6.2 ACK-based Fingerprinting 6.2 ACK-based Fingerprinting 6.3 Active Server Fingerprinting 6.3 Active Server Fingerprinting 6.4 Constructing Filters and Probers 6.4 Constructing Filters and Probers 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7.2 Choice of Observation Window N 7.2 Choice of Observation Window N 7.3 Effects of Packet Loss 7.3 Effects of Packet Loss 7.4 Server Churn for Asynchronous Probing 7.4 Server Churn for Asynchronous Probing 7.5 Probe UDP and Obfuscated OpenVPN Servers 7.5 Probe UDP and Obfuscated OpenVPN Servers 8 Real-world Deployment Setup 8 Real-world Deployment Setup 9 Evaluation & Findings and 9.1 Results for control VPN flows 9 Evaluation & Findings and 9.1 Results for control VPN flows 9.2 Results for all flows 9.2 Results for all flows 10 Discussion and Mitigations 10 Discussion and Mitigations 11 Conclusion 11 Conclusion 12 Acknowledgement and References 12 Acknowledgement and References Appendix Appendix A Appendix This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

OpenVPN is Open to VPN Fingerprinting: Appendix

OpenVPN is Open to VPN Fingerprinting: Acknowledgement and References

Authors:
(1) Diwen Xue, University of Michigan;
(2) Reethika Ramesh, University of Michigan;
(3) Arham Jain, University of Michigan;
(4) Arham Jain, Merit Network, Inc.;
(5) J. Alex Halderman, University of Michigan;
(6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad;
(7) Roya Ensaf, University of Michigan. Table of Links Abstract and 1 Introduction 2 Background & Related Work 3 Challenges in Real-world VPN Detection 4 Adversary Model and Deployment 5 Ethics, Privacy, and Responsible Disclosure 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6.2 ACK-based Fingerprinting 6.3 Active Server Fingerprinting 6.4 Constructing Filters and Probers 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7.2 Choice of Observation Window N 7.3 Effects of Packet Loss 7.4 Server Churn for Asynchronous Probing 7.5 Probe UDP and Obfuscated OpenVPN Servers 8 Real-world Deployment Setup 9 Evaluation & Findings and 9.1 Results for control VPN flows 9.2 Results for all flows 10 Discussion and Mitigations 11 Conclusion 12 Acknowledgement and References Appendix 10 Discussion and Mitigations ISPs and government censors are motivated to detect OpenVPN flows in order to enforce traffic policies and information controls. We demonstrate that tracking and blocking the use of OpenVPN, even with most deployed obfuscation methods, is practical at scale and with minimal collateral damage. We note that many VPN providers’ claims that their obfuscated services are unobservable appear to be misleading and potentially dangerous, especially to users from countries where personal VPN usage is illegal. In light of our findings, users should not expect complete unobservability even when connected to “obfuscated” OpenVPN-based services. Putting the human danger aside, the ease of fingerprinting makes OpenVPN more susceptible to throttling or blocking from ISPs and governments. Previous research suggests that some censors already use two-stage pipelines, which are highly similar to our deployment, to detect other protocols such as Tor or Shadowsocks [1, 71]. These adversaries can quickly adapt such infrastructure to detect OpenVPN traffic by simply adding protocol-specific fingerprints and probes. Furthermore, while we focus on OpenVPN due to its overwhelming popularity among commercial VPNs, it is possible to extend our two-stage framework to other VPN protocols (e.g., WireGuard and StrongSwan) by analyzing their communication patterns and server behaviors. Governments can also quickly adopt these fingerprints to track and block VPN usage during sensitive times, like political upheavals, when VPN connections are most vital to the free flow of information. Short-Term Mitigation There are several defensive strategies to achieve near-term protection from the fingerprinting attack we describe. First, VPN providers offering both vanilla and obfuscated OpenVPN services should avoid colocating them. Ideally, obfuscation servers should be well separated from OpenVPN instances in the network address space and operate as “bridge servers” that forward client traffic to VPN servers elsewhere. For example, Mullvad VPN offers a Shadowsocks-based obfuscator service as a dedicated bridge, separating the VPN servers from the obfuscation [28]. Second, VPN providers should switch from static to random padding for their obfuscated services. As we have shown, for protocols with a stable and distinctive handshake phase, even the most basic threshold-based detector is able to fingerprint them by packet sizes. Ideally, the obfuscation layer should be able to send zero-length packets (packets whose payloads are all padding) to break the one-to-one correspondence between the unobfuscated and obfuscated packet streams [72]. Yet it is worth noting that previous work has shown that even fully randomized obfuscators (e.g., obfs4) can themselves be vulnerable to entropy-based fingerprinting attacks [67]. Third, we suggest that the OpenVPN developers follow recommendations from previous work with regard to how servers respond to failed handshake attempts. Servers closing failed connections immediately or in a predictable manner has enabled active probing attacks against a variety of other protocols [1, 13, 58]. In response, these protocols have implemented either unlimited timeouts (reading from the buffer indefinitely) or diversified close behaviors (in which each server instance closes failed connections in a different manner). Long-Term Defenses In the long term, we fear that the catand-mouse game between censors and circumvention tools, such as the Great Firewall and Tor, will occur in the VPN ecosystem as well, and developers and providers will have to adapt their obfuscation strategies to the evolving adversaries. We urge commercial VPN providers to adopt more standardized obfuscation solutions, such as Pluggable Transports [39], and to be more transparent about the techniques used by their obfuscated services. This transparency will help foster development of stronger obfuscation methods and encourage developers to design better techniques to overcome the progress of information control technologies. Additional future work is needed to characterize the performance costs of different approaches to VPN obfuscation and to help users with varying threat models make appropriate trade-offs between performance and resilient unobservability. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Diwen Xue, University of Michigan; (2) Reethika Ramesh, University of Michigan; (3) Arham Jain, University of Michigan; (4) Arham Jain, Merit Network, Inc.; (5) J. Alex Halderman, University of Michigan; (6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad; (7) Roya Ensaf, University of Michigan. Authors: Authors: (1) Diwen Xue, University of Michigan; (2) Reethika Ramesh, University of Michigan; (3) Arham Jain, University of Michigan; (4) Arham Jain, Merit Network, Inc.; (5) J. Alex Halderman, University of Michigan; (6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad; (7) Roya Ensaf, University of Michigan. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Background & Related Work 2 Background & Related Work 3 Challenges in Real-world VPN Detection 3 Challenges in Real-world VPN Detection 4 Adversary Model and Deployment 4 Adversary Model and Deployment 5 Ethics, Privacy, and Responsible Disclosure 5 Ethics, Privacy, and Responsible Disclosure 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6.2 ACK-based Fingerprinting 6.2 ACK-based Fingerprinting 6.3 Active Server Fingerprinting 6.3 Active Server Fingerprinting 6.4 Constructing Filters and Probers 6.4 Constructing Filters and Probers 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7.2 Choice of Observation Window N 7.2 Choice of Observation Window N 7.3 Effects of Packet Loss 7.3 Effects of Packet Loss 7.4 Server Churn for Asynchronous Probing 7.4 Server Churn for Asynchronous Probing 7.5 Probe UDP and Obfuscated OpenVPN Servers 7.5 Probe UDP and Obfuscated OpenVPN Servers 8 Real-world Deployment Setup 8 Real-world Deployment Setup 9 Evaluation & Findings and 9.1 Results for control VPN flows 9 Evaluation & Findings and 9.1 Results for control VPN flows 9.2 Results for all flows 9.2 Results for all flows 10 Discussion and Mitigations 10 Discussion and Mitigations 11 Conclusion 11 Conclusion 12 Acknowledgement and References 12 Acknowledgement and References Appendix Appendix 10 Discussion and Mitigations ISPs and government censors are motivated to detect OpenVPN flows in order to enforce traffic policies and information controls. We demonstrate that tracking and blocking the use of OpenVPN, even with most deployed obfuscation methods, is practical at scale and with minimal collateral damage. We note that many VPN providers’ claims that their obfuscated services are unobservable appear to be misleading and potentially dangerous, especially to users from countries where personal VPN usage is illegal. In light of our findings, users should not expect complete unobservability even when connected to “obfuscated” OpenVPN-based services. Putting the human danger aside, the ease of fingerprinting makes OpenVPN more susceptible to throttling or blocking from ISPs and governments. Previous research suggests that some censors already use two-stage pipelines, which are highly similar to our deployment, to detect other protocols such as Tor or Shadowsocks [1, 71]. These adversaries can quickly adapt such infrastructure to detect OpenVPN traffic by simply adding protocol-specific fingerprints and probes. Furthermore, while we focus on OpenVPN due to its overwhelming popularity among commercial VPNs, it is possible to extend our two-stage framework to other VPN protocols (e.g., WireGuard and StrongSwan) by analyzing their communication patterns and server behaviors. Governments can also quickly adopt these fingerprints to track and block VPN usage during sensitive times, like political upheavals, when VPN connections are most vital to the free flow of information. Short-Term Mitigation There are several defensive strategies to achieve near-term protection from the fingerprinting attack we describe. First, VPN providers offering both vanilla and obfuscated OpenVPN services should avoid colocating them. Ideally, obfuscation servers should be well separated from OpenVPN instances in the network address space and operate as “bridge servers” that forward client traffic to VPN servers elsewhere. For example, Mullvad VPN offers a Shadowsocks-based obfuscator service as a dedicated bridge, separating the VPN servers from the obfuscation [28]. Short-Term Mitigation Second, VPN providers should switch from static to random padding for their obfuscated services. As we have shown, for protocols with a stable and distinctive handshake phase, even the most basic threshold-based detector is able to fingerprint them by packet sizes. Ideally, the obfuscation layer should be able to send zero-length packets (packets whose payloads are all padding) to break the one-to-one correspondence between the unobfuscated and obfuscated packet streams [72]. Yet it is worth noting that previous work has shown that even fully randomized obfuscators (e.g., obfs4) can themselves be vulnerable to entropy-based fingerprinting attacks [67]. Third, we suggest that the OpenVPN developers follow recommendations from previous work with regard to how servers respond to failed handshake attempts. Servers closing failed connections immediately or in a predictable manner has enabled active probing attacks against a variety of other protocols [1, 13, 58]. In response, these protocols have implemented either unlimited timeouts (reading from the buffer indefinitely) or diversified close behaviors (in which each server instance closes failed connections in a different manner). Long-Term Defenses In the long term, we fear that the catand-mouse game between censors and circumvention tools, such as the Great Firewall and Tor, will occur in the VPN ecosystem as well, and developers and providers will have to adapt their obfuscation strategies to the evolving adversaries. We urge commercial VPN providers to adopt more standardized obfuscation solutions, such as Pluggable Transports [39], and to be more transparent about the techniques used by their obfuscated services. This transparency will help foster development of stronger obfuscation methods and encourage developers to design better techniques to overcome the progress of information control technologies. Additional future work is needed to characterize the performance costs of different approaches to VPN obfuscation and to help users with varying threat models make appropriate trade-offs between performance and resilient unobservability. Long-Term Defenses This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Mitigating OpenVPN Fingerprinting: Risks and Defensive Strategies

Authors:
(1) Diwen Xue, University of Michigan;
(2) Reethika Ramesh, University of Michigan;
(3) Arham Jain, University of Michigan;
(4) Arham Jain, Merit Network, Inc.;
(5) J. Alex Halderman, University of Michigan;
(6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad;
(7) Roya Ensaf, University of Michigan. Table of Links Abstract and 1 Introduction 2 Background & Related Work 3 Challenges in Real-world VPN Detection 4 Adversary Model and Deployment 5 Ethics, Privacy, and Responsible Disclosure 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6.2 ACK-based Fingerprinting 6.3 Active Server Fingerprinting 6.4 Constructing Filters and Probers 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7.2 Choice of Observation Window N 7.3 Effects of Packet Loss 7.4 Server Churn for Asynchronous Probing 7.5 Probe UDP and Obfuscated OpenVPN Servers 8 Real-world Deployment Setup 9 Evaluation & Findings and 9.1 Results for control VPN flows 9.2 Results for all flows 10 Discussion and Mitigations 11 Conclusion 12 Acknowledgement and References Appendix 8 Real-world Deployment Setup We set out to explore if an ISP or censor can fingerprint OpenVPN connections at scale, without significant collateral damage. Adopting the viewpoint of an adversarial ISP, we deploy our framework inside Merit, as shown in Figure 2. Our evaluation is two-fold: we generate control vanilla and obfuscated flows with commercial VPN providers and attempt to identify them as a network intermediary; we also process other traffic passing through our Monitoring Station in order to estimate the false positive rate of our framework. We set up our framework on a 16-core server (Monitoring Station) inside Merit with two mirroring interfaces that have an aggregated 20 Gbps bandwidth. Due to the large traffic volume, we optimize our deployment with PF_RING [38] in order to improve the packet processing speed. We employ PF_- RING in zero-copy mode and spread the traffic load across a Zeek cluster of 15 workers. Nonetheless, due to limited CPU resources, we only sample 12.5% of all TCP and UDP flows arriving at the network interfaces in order to minimize the effect of packet loss. The sampling is based on IP pairs so that all bi-directional traffic of a flow will be selected/dropped together. With these settings, we are able to operate with an end-to-end packet loss rate under 3%. Even though we process only a fraction of all traffic, our Filter still handles over 15 Terabytes of traffic from over 2 billion flows on an average day on a single server. In addition, processing all traffic without sampling is feasible through parallelism or using faster CPUs. Next, we set up Probers on two dedicated measurement machines, each provisioned with 10 IPv4 and 1 IPv6 addresses. By the end of each day during the evaluation, the Probers fetch filtering logs from the Monitoring Station. For each target, we run a Masscan [25] to the /29 subnet the IP belongs to over all TCP ports (1-65535). We follow up each discovered open port by running our probing scheme, and endpoints confirmed through probing are recorded for manual analysis. To select VPN services for evaluation, we first generate a list of “top” VPN services ranked by popularity. We combine 80 providers, most of which are paid premium VPN services, from top VPN recommendation sites based on previous work [42], listed in Appendix Table 4. Next, we visit the websites of these VPN providers searching for “Obfuscation”, “Stealth”, or “Camouflage Mode” etc., and include providers that offer at least one obfuscated VPN configuration. In total, we find 24 providers offering obfuscated services. We test all obfuscation configurations if more than one is offered as well as vanilla OpenVPN for each provider. If TCP and UDP modes are both available, we test them separately. In total, we have 81 configurations, 41 of which are obfuscated ones. We configure the Client Station inside Merit to act as a VPN client. Both upstream and downstream traffic of the Client Station go through the router that mirrors traffic to the Monitoring Station. In addition, we exclude this server from our random sampling so that all traffic to/from this server will be analyzed. On the client, we run an automated script to generate control traffic for our evaluation. For each iteration, we start the VPN client application and connect to the “default / recommended” server using Pywinauto [41]. After a random wait of 20 to 180 seconds, we confirm that the VPN tunnel is active and generate random browsing traffic with Selenium [45] by sending requests to a random website from the Alexa top 500. Finally, we disconnect from the VPN server and wait for 180 seconds before proceeding to the next iteration. For each VPN configuration, we repeat the process 50 times and collect packet captures for reference. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Diwen Xue, University of Michigan; (2) Reethika Ramesh, University of Michigan; (3) Arham Jain, University of Michigan; (4) Arham Jain, Merit Network, Inc.; (5) J. Alex Halderman, University of Michigan; (6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad; (7) Roya Ensaf, University of Michigan. Authors: Authors: (1) Diwen Xue, University of Michigan; (2) Reethika Ramesh, University of Michigan; (3) Arham Jain, University of Michigan; (4) Arham Jain, Merit Network, Inc.; (5) J. Alex Halderman, University of Michigan; (6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad; (7) Roya Ensaf, University of Michigan. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Background & Related Work 2 Background & Related Work 3 Challenges in Real-world VPN Detection 3 Challenges in Real-world VPN Detection 4 Adversary Model and Deployment 4 Adversary Model and Deployment 5 Ethics, Privacy, and Responsible Disclosure 5 Ethics, Privacy, and Responsible Disclosure 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6.2 ACK-based Fingerprinting 6.2 ACK-based Fingerprinting 6.3 Active Server Fingerprinting 6.3 Active Server Fingerprinting 6.4 Constructing Filters and Probers 6.4 Constructing Filters and Probers 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7.2 Choice of Observation Window N 7.2 Choice of Observation Window N 7.3 Effects of Packet Loss 7.3 Effects of Packet Loss 7.4 Server Churn for Asynchronous Probing 7.4 Server Churn for Asynchronous Probing 7.5 Probe UDP and Obfuscated OpenVPN Servers 7.5 Probe UDP and Obfuscated OpenVPN Servers 8 Real-world Deployment Setup 8 Real-world Deployment Setup 9 Evaluation & Findings and 9.1 Results for control VPN flows 9 Evaluation & Findings and 9.1 Results for control VPN flows 9.2 Results for all flows 9.2 Results for all flows 10 Discussion and Mitigations 10 Discussion and Mitigations 11 Conclusion 11 Conclusion 12 Acknowledgement and References 12 Acknowledgement and References Appendix Appendix 8 Real-world Deployment Setup We set out to explore if an ISP or censor can fingerprint OpenVPN connections at scale, without significant collateral damage. Adopting the viewpoint of an adversarial ISP, we deploy our framework inside Merit, as shown in Figure 2. Our evaluation is two-fold: we generate control vanilla and obfuscated flows with commercial VPN providers and attempt to identify them as a network intermediary; we also process other traffic passing through our Monitoring Station in order to estimate the false positive rate of our framework. We set up our framework on a 16-core server (Monitoring Station) inside Merit with two mirroring interfaces that have an aggregated 20 Gbps bandwidth. Due to the large traffic volume, we optimize our deployment with PF_RING [38] in order to improve the packet processing speed. We employ PF_- RING in zero-copy mode and spread the traffic load across a Zeek cluster of 15 workers. Nonetheless, due to limited CPU resources, we only sample 12.5% of all TCP and UDP flows arriving at the network interfaces in order to minimize the effect of packet loss. The sampling is based on IP pairs so that all bi-directional traffic of a flow will be selected/dropped together. With these settings, we are able to operate with an end-to-end packet loss rate under 3%. Even though we process only a fraction of all traffic, our Filter still handles over 15 Terabytes of traffic from over 2 billion flows on an average day on a single server. In addition, processing all traffic without sampling is feasible through parallelism or using faster CPUs. Next, we set up Probers on two dedicated measurement machines, each provisioned with 10 IPv4 and 1 IPv6 addresses. By the end of each day during the evaluation, the Probers fetch filtering logs from the Monitoring Station. For each target, we run a Masscan [25] to the /29 subnet the IP belongs to over all TCP ports (1-65535). We follow up each discovered open port by running our probing scheme, and endpoints confirmed through probing are recorded for manual analysis. To select VPN services for evaluation, we first generate a list of “top” VPN services ranked by popularity. We combine 80 providers, most of which are paid premium VPN services, from top VPN recommendation sites based on previous work [42], listed in Appendix Table 4. Next, we visit the websites of these VPN providers searching for “Obfuscation”, “Stealth”, or “Camouflage Mode” etc., and include providers that offer at least one obfuscated VPN configuration. In total, we find 24 providers offering obfuscated services. We test all obfuscation configurations if more than one is offered as well as vanilla OpenVPN for each provider. If TCP and UDP modes are both available, we test them separately. In total, we have 81 configurations, 41 of which are obfuscated ones. We configure the Client Station inside Merit to act as a VPN client. Both upstream and downstream traffic of the Client Station go through the router that mirrors traffic to the Monitoring Station. In addition, we exclude this server from our random sampling so that all traffic to/from this server will be analyzed. On the client, we run an automated script to generate control traffic for our evaluation. For each iteration, we start the VPN client application and connect to the “default / recommended” server using Pywinauto [41]. After a random wait of 20 to 180 seconds, we confirm that the VPN tunnel is active and generate random browsing traffic with Selenium [45] by sending requests to a random website from the Alexa top 500. Finally, we disconnect from the VPN server and wait for 180 seconds before proceeding to the next iteration. For each VPN configuration, we repeat the process 50 times and collect packet captures for reference. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

OpenVPN is Open to VPN Fingerprinting: Real-world Deployment Setup

Authors:
(1) Diwen Xue, University of Michigan;
(2) Reethika Ramesh, University of Michigan;
(3) Arham Jain, University of Michigan;
(4) Arham Jain, Merit Network, Inc.;
(5) J. Alex Halderman, University of Michigan;
(6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad;
(7) Roya Ensaf, University of Michigan. Table of Links Abstract and 1 Introduction 2 Background & Related Work 3 Challenges in Real-world VPN Detection 4 Adversary Model and Deployment 5 Ethics, Privacy, and Responsible Disclosure 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6.2 ACK-based Fingerprinting 6.3 Active Server Fingerprinting 6.4 Constructing Filters and Probers 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7.2 Choice of Observation Window N 7.3 Effects of Packet Loss 7.4 Server Churn for Asynchronous Probing 7.5 Probe UDP and Obfuscated OpenVPN Servers 8 Real-world Deployment Setup 9 Evaluation & Findings and 9.1 Results for control VPN flows 9.2 Results for all flows 10 Discussion and Mitigations 11 Conclusion 12 Acknowledgement and References Appendix 6.4 Constructing Filters and Probers Our Filter performs both opcode and ACK-based fingerprinting, flagging a flow if at least one fingerprint matches. This is because the opcode and ACK fingerprints are designed to be complementary: both are effective against vanilla OpenVPN and they each target a specific subset of obfuscations. The former works against XOR-based obfuscations that work like Vigenère ciphers, i.e. they always encrypt the same plaintext opcodes at the same position to the same ciphertext bytes. The latter targets tunneling-based obfuscation that lacks random padding and preserves the 1:1 correspondence between the original and obfuscated packet streams. Combining the two features maximizes our fingerprinting coverage, as we discovered that even within the same provider, obfuscating strategies can vary a lot (§ 9). Table 5 in Appendix shows the effectiveness of each feature against each commercial VPN service we tested. Following Filter’s result, the Prober performs the active probing scheme to further lower potential false positives. We implement the Filter in Zeek [75], an open-source network monitoring tool. We note that the evaluation processes for opcode and ACK-based fingerprinting are quite simple: both only require several dozen integer comparisons (limited by the observation window) while maintaining a small number of per-flow states. We implement the Prober in Nim [31]. We believe that both components can be easily deployed by any ISP or censor. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Diwen Xue, University of Michigan; (2) Reethika Ramesh, University of Michigan; (3) Arham Jain, University of Michigan; (4) Arham Jain, Merit Network, Inc.; (5) J. Alex Halderman, University of Michigan; (6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad; (7) Roya Ensaf, University of Michigan. Authors: Authors: (1) Diwen Xue, University of Michigan; (2) Reethika Ramesh, University of Michigan; (3) Arham Jain, University of Michigan; (4) Arham Jain, Merit Network, Inc.; (5) J. Alex Halderman, University of Michigan; (6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad; (7) Roya Ensaf, University of Michigan. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Background & Related Work 2 Background & Related Work 3 Challenges in Real-world VPN Detection 3 Challenges in Real-world VPN Detection 4 Adversary Model and Deployment 4 Adversary Model and Deployment 5 Ethics, Privacy, and Responsible Disclosure 5 Ethics, Privacy, and Responsible Disclosure 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6.2 ACK-based Fingerprinting 6.2 ACK-based Fingerprinting 6.3 Active Server Fingerprinting 6.3 Active Server Fingerprinting 6.4 Constructing Filters and Probers 6.4 Constructing Filters and Probers 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7.2 Choice of Observation Window N 7.2 Choice of Observation Window N 7.3 Effects of Packet Loss 7.3 Effects of Packet Loss 7.4 Server Churn for Asynchronous Probing 7.4 Server Churn for Asynchronous Probing 7.5 Probe UDP and Obfuscated OpenVPN Servers 7.5 Probe UDP and Obfuscated OpenVPN Servers 8 Real-world Deployment Setup 8 Real-world Deployment Setup 9 Evaluation & Findings and 9.1 Results for control VPN flows 9 Evaluation & Findings and 9.1 Results for control VPN flows 9.2 Results for all flows 9.2 Results for all flows 10 Discussion and Mitigations 10 Discussion and Mitigations 11 Conclusion 11 Conclusion 12 Acknowledgement and References 12 Acknowledgement and References Appendix Appendix 6.4 Constructing Filters and Probers Filters Our Filter performs both opcode and ACK-based fingerprinting, flagging a flow if at least one fingerprint matches. This is because the opcode and ACK fingerprints are designed to be complementary: both are effective against vanilla OpenVPN and they each target a specific subset of obfuscations. The former works against XOR-based obfuscations that work like Vigenère ciphers, i.e. they always encrypt the same plaintext opcodes at the same position to the same ciphertext bytes. The latter targets tunneling-based obfuscation that lacks random padding and preserves the 1:1 correspondence between the original and obfuscated packet streams. Combining the two features maximizes our fingerprinting coverage, as we discovered that even within the same provider, obfuscating strategies can vary a lot (§ 9). Table 5 in Appendix shows the effectiveness of each feature against each commercial VPN service we tested. Following Filter’s result, the Prober performs the active probing scheme to further lower potential false positives. Filter Filter’s Prober We implement the Filter in Zeek [75], an open-source network monitoring tool. We note that the evaluation processes for opcode and ACK-based fingerprinting are quite simple: both only require several dozen integer comparisons (limited by the observation window) while maintaining a small number of per-flow states. We implement the Prober in Nim [31]. We believe that both components can be easily deployed by any ISP or censor. Filter Prober This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Enhancing OpenVPN Detection: Combining Fingerprinting and Active Probing

Authors:
(1) Diwen Xue, University of Michigan;
(2) Reethika Ramesh, University of Michigan;
(3) Arham Jain, University of Michigan;
(4) Arham Jain, Merit Network, Inc.;
(5) J. Alex Halderman, University of Michigan;
(6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad;
(7) Roya Ensaf, University of Michigan. Table of Links Abstract and 1 Introduction 2 Background & Related Work 3 Challenges in Real-world VPN Detection 4 Adversary Model and Deployment 5 Ethics, Privacy, and Responsible Disclosure 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6.2 ACK-based Fingerprinting 6.3 Active Server Fingerprinting 6.4 Constructing Filters and Probers 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7.2 Choice of Observation Window N 7.3 Effects of Packet Loss 7.4 Server Churn for Asynchronous Probing 7.5 Probe UDP and Obfuscated OpenVPN Servers 8 Real-world Deployment Setup 9 Evaluation & Findings and 9.1 Results for control VPN flows 9.2 Results for all flows 10 Discussion and Mitigations 11 Conclusion 12 Acknowledgement and References Appendix 6.3 Active Server Fingerprinting We explore the feasibility of identifying an OpenVPN server through active probing. Typically, OpenVPN servers respond to a client reset with an explicit server reset, thereby giving away their identity. However, most commercial providers now have adopted tls-auth or tls-crypt options [50]. These options add an additional HMAC signature—signed by a pre-shared key—to every control channel packet for integrity verification, including the initial reset packets. With either of these options enabled, an OpenVPN server would not respond to an unauthenticated client reset with a server reset, but would instead drop such packets without further processing. The presence of such HMAC mechanism increases the complexity of doing active probing: it effectively makes OpenVPN servers “probe-resistant” [13] by remaining silent when probed by an unauthenticated client. In fact, similar HMAC mechanisms are used by more popular “probe-resistant” proxies, such as obfs4 [33]. However, unlike obfs4 which waits for a server-specific random delay before dropping an unauthenticated connection, OpenVPN always immediately closes the connection if a valid HMAC cannot be located. We design our probes to leverage this protocol-specific behavior, and as a result, we manage to fingerprint OpenVPN servers even if they do not respond throughout our probing cycles. The key concept is that although the application may not respond to probing, an attacker may still be able to fingerprint application-specific thresholds at the TCP level, such as timeouts or RST thresholds, as demonstrated by Frolov et al. [13]. We use two datasets in this section to help with designing probes. ZMap Set: to construct a realistic non-VPN endpoints dataset, we use ZMap to scan each of the 65,535 TCP ports over the entire IPv4 space, limiting results for each port to 200 endpoints (with the specific port open), resulting in over 13 million endpoints. Censys Set: We query the Censys.io [10] database for hosts with TCP port 1194/OpenVPN open. Next, we probe each endpoint with a typical OpenVPN Client Reset and group endpoints that respond with explicit Server Resets. This results in 180,858 hosts known to be OpenVPN endpoints (with “tls-auth” disabled). 6.3.1 Base Probes We design probes exploiting a behavior associated with how OpenVPN packetizes TCP streams. When OpenVPN operates over TCP, it needs to split the continuous stream into discrete OpenVPN packets. Figure 5 presents a high-level abstraction of this process. The most relevant parts are: a buffer is allocated in memory to reassemble fragments of OpenVPN packets encapsulated in TCP streams. The length N for the next OpenVPN packet is extracted from the first two bytes of the header (see Figure 3), and the routine keeps reading N additional bytes before it returns the reassembled packet to the caller. This means that an OpenVPN packet will not be parsed and checked for syntax and encryption errors until all its parts arrive at the server. Based on this behavior, we design two sequential probes to trigger an OpenVPN server into different code paths—which result in different connection timeouts— and measure the time elapsed before the server responds or terminates the connection. As shown in Table 1, Base Probe 1 carries a typical 16-byte OpenVPN Client Reset, while Base Probe 2 has the same payload with the last byte stripped off. The assumption is since our two probes only differ in one byte, most non-OpenVPN servers will respond to our probes in a similar way. However, for an OpenVPN server with HMAC enabled, the connection sending the first probe will be dropped immediately because the OpenVPN packet is reassembled and a valid HMAC cannot be located. The second probe will not receive an immediate response, as the server will wait for an additional byte to arrive for reassembly. The connection will stay idle until a server specific handshake timeout has passed, after which the connection will be dropped. As such, the first probe will be dropped at the decryption routine, while the second probe will be dropped at the packet reassembly routine, both labeled red in Figure 5. 6.3.2 Additional Probes The two probes, although useful, are limited and there may be other protocols with behaviors similar to OpenVPN. After using both to probe the ZMap Set, we still identify a handful of services that respond similarly to OpenVPN servers, such as Microsoft WBT Server (3389), Microsoft Message Queuing (1801), and Erlang Port Mapper Daemon (4369). We design additional probes based on the fact that OpenVPN validates packet length and will drop connections sending invalid length without waiting for the next packet to be reassembled. Here, packet length refers to the length declared by the first two bytes of an OpenVPN header (see Figure 3), rather than the TCP packet length. A “valid” length is in the range of [1, max_len], where max_len is derived from the server’s MTU configurations. For instance, default TUN MTU of 1500 bytes, combined with overheads (crypto IV, packet length, etc.), results in a max_len of 1627 bytes. In this case, probes whose first two bytes have a decimal value greater than 1627 (0x06,0x5B) will be dropped immediately. We also design probes leveraging the way a Linux server closes a TCP connection. When a TCP connection terminates, the operating systems at both ends typically complete a FIN 4-way handshake. However, previous work has found that if a connection is closed with unread bytes in buffer, Linux will send a RST packet [13]. A server’s “RST Threshold” is defined as the minimum number of bytes needed to send to the server to trigger a RST. We determine the RST threshold distribution for both ZMap Set and Censys Set. As shown in Figure 6, the vast majority of OpenVPN servers have a RST threshold around 1550-1660 bytes, corresponding to buffers allocated with typical MTU configurations. In contrast, over 97% of random ZMap endpoints have a RST threshold less than 500 or greater than 4000. We therefore construct an additional probe with 2,000 random bytes, which we expect over 98% of legitimate OpenVPN servers and less than 3% of random servers to respond to with RST packets. It is worth noting that the majority of “typical” HTTP, TLS, and SSH servers have already been filtered out by our base probes, so endpoints that respond at this stage are likely sharing the port with another service, thus warranting manual analysis (e.g., checking TLS certificate). While these three services are what we commonly observed, there may be instances where other services are running along with OpenVPN. This could lead to false negatives. Table 1 lists all probes and the expected behaviors from an OpenVPN server. An evaluation process is shown in Appendix Figure 11. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Diwen Xue, University of Michigan; (2) Reethika Ramesh, University of Michigan; (3) Arham Jain, University of Michigan; (4) Arham Jain, Merit Network, Inc.; (5) J. Alex Halderman, University of Michigan; (6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad; (7) Roya Ensaf, University of Michigan. Authors: Authors: (1) Diwen Xue, University of Michigan; (2) Reethika Ramesh, University of Michigan; (3) Arham Jain, University of Michigan; (4) Arham Jain, Merit Network, Inc.; (5) J. Alex Halderman, University of Michigan; (6) Jedidiah R. Crandall, Arizona State University/Breakpointing Bad; (7) Roya Ensaf, University of Michigan. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Background & Related Work 2 Background & Related Work 3 Challenges in Real-world VPN Detection 3 Challenges in Real-world VPN Detection 4 Adversary Model and Deployment 4 Adversary Model and Deployment 5 Ethics, Privacy, and Responsible Disclosure 5 Ethics, Privacy, and Responsible Disclosure 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6 Identifying Fingerprintable Features and 6.1 Opcode-based Fingerprinting 6.2 ACK-based Fingerprinting 6.2 ACK-based Fingerprinting 6.3 Active Server Fingerprinting 6.3 Active Server Fingerprinting 6.4 Constructing Filters and Probers 6.4 Constructing Filters and Probers 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7 Fine-tuning for Deployment and 7.1 ACK Fingerprint Thresholds 7.2 Choice of Observation Window N 7.2 Choice of Observation Window N 7.3 Effects of Packet Loss 7.3 Effects of Packet Loss 7.4 Server Churn for Asynchronous Probing 7.4 Server Churn for Asynchronous Probing 7.5 Probe UDP and Obfuscated OpenVPN Servers 7.5 Probe UDP and Obfuscated OpenVPN Servers 8 Real-world Deployment Setup 8 Real-world Deployment Setup 9 Evaluation & Findings and 9.1 Results for control VPN flows 9 Evaluation & Findings and 9.1 Results for control VPN flows 9.2 Results for all flows 9.2 Results for all flows 10 Discussion and Mitigations 10 Discussion and Mitigations 11 Conclusion 11 Conclusion 12 Acknowledgement and References 12 Acknowledgement and References Appendix Appendix 6.3 Active Server Fingerprinting We explore the feasibility of identifying an OpenVPN server through active probing. Typically, OpenVPN servers respond to a client reset with an explicit server reset, thereby giving away their identity. However, most commercial providers now have adopted tls-auth or tls-crypt options [50]. These options add an additional HMAC signature—signed by a pre-shared key—to every control channel packet for integrity verification, including the initial reset packets. With either of these options enabled, an OpenVPN server would not respond to an unauthenticated client reset with a server reset, but would instead drop such packets without further processing. The presence of such HMAC mechanism increases the complexity of doing active probing: it effectively makes OpenVPN servers “probe-resistant” [13] by remaining silent when probed by an unauthenticated client. In fact, similar HMAC mechanisms are used by more popular “probe-resistant” proxies, such as obfs4 [33]. However, unlike obfs4 which waits for a server-specific random delay before dropping an unauthenticated connection, OpenVPN always immediately closes the connection if a valid HMAC cannot be located. We design our probes to leverage this protocol-specific behavior, and as a result, we manage to fingerprint OpenVPN servers even if they do not respond throughout our probing cycles. The key concept is that although the application may not respond to probing, an attacker may still be able to fingerprint application-specific thresholds at the TCP level, such as timeouts or RST thresholds, as demonstrated by Frolov et al. [13]. We use two datasets in this section to help with designing probes. ZMap Set: to construct a realistic non-VPN endpoints dataset, we use ZMap to scan each of the 65,535 TCP ports over the entire IPv4 space, limiting results for each port to 200 endpoints (with the specific port open), resulting in over 13 million endpoints. Censys Set: We query the Censys.io [10] database for hosts with TCP port 1194/OpenVPN open. Next, we probe each endpoint with a typical OpenVPN Client Reset and group endpoints that respond with explicit Server Resets. This results in 180,858 hosts known to be OpenVPN endpoints (with “tls-auth” disabled). 6.3.1 Base Probes 6.3.1 Base Probes We design probes exploiting a behavior associated with how OpenVPN packetizes TCP streams. When OpenVPN operates over TCP, it needs to split the continuous stream into discrete OpenVPN packets. Figure 5 presents a high-level abstraction of this process. The most relevant parts are: a buffer is allocated in memory to reassemble fragments of OpenVPN packets encapsulated in TCP streams. The length N for the next OpenVPN packet is extracted from the first two bytes of the header (see Figure 3), and the routine keeps reading N additional bytes before it returns the reassembled packet to the caller. This means that an OpenVPN packet will not be parsed and checked for syntax and encryption errors until all its parts arrive at the server. Based on this behavior, we design two sequential probes to trigger an OpenVPN server into different code paths—which result in different connection timeouts— and measure the time elapsed before the server responds or terminates the connection. As shown in Table 1, Base Probe 1 carries a typical 16-byte OpenVPN Client Reset, while Base Probe 2 has the same payload with the last byte stripped off. The assumption is since our two probes only differ in one byte, most non-OpenVPN servers will respond to our probes in a similar way. However, for an OpenVPN server with HMAC enabled, the connection sending the first probe will be dropped immediately because the OpenVPN packet is reassembled and a valid HMAC cannot be located. The second probe will not receive an immediate response, as the server will wait for an additional byte to arrive for reassembly. The connection will stay idle until a server specific handshake timeout has passed, after which the connection will be dropped. As such, the first probe will be dropped at the decryption routine, while the second probe will be dropped at the packet reassembly routine, both labeled red in Figure 5. 6.3.2 Additional Probes 6.3.2 Additional Probes The two probes, although useful, are limited and there may be other protocols with behaviors similar to OpenVPN. After using both to probe the ZMap Set, we still identify a handful of services that respond similarly to OpenVPN servers, such as Microsoft WBT Server (3389), Microsoft Message Queuing (1801), and Erlang Port Mapper Daemon (4369). We design additional probes based on the fact that OpenVPN validates packet length and will drop connections sending invalid length without waiting for the next packet to be reassembled. Here, packet length refers to the length declared by the first two bytes of an OpenVPN header (see Figure 3), rather than the TCP packet length. A “valid” length is in the range of [1, max_len], where max_len is derived from the server’s MTU configurations. For instance, default TUN MTU of 1500 bytes, combined with overheads (crypto IV, packet length, etc.), results in a max_len of 1627 bytes. In this case, probes whose first two bytes have a decimal value greater than 1627 (0x06,0x5B) will be dropped immediately. We also design probes leveraging the way a Linux server closes a TCP connection. When a TCP connection terminates, the operating systems at both ends typically complete a FIN 4-way handshake. However, previous work has found that if a connection is closed with unread bytes in buffer, Linux will send a RST packet [13]. A server’s “RST Threshold” is defined as the minimum number of bytes needed to send to the server to trigger a RST. We determine the RST threshold distribution for both ZMap Set and Censys Set. As shown in Figure 6, the vast majority of OpenVPN servers have a RST threshold around 1550-1660 bytes, corresponding to buffers allocated with typical MTU configurations. In contrast, over 97% of random ZMap endpoints have a RST threshold less than 500 or greater than 4000. We therefore construct an additional probe with 2,000 random bytes, which we expect over 98% of legitimate OpenVPN servers and less than 3% of random servers to respond to with RST packets. It is worth noting that the majority of “typical” HTTP, TLS, and SSH servers have already been filtered out by our base probes, so endpoints that respond at this stage are likely sharing the port with another service, thus warranting manual analysis (e.g., checking TLS certificate). While these three services are what we commonly observed, there may be instances where other services are running along with OpenVPN. This could lead to false negatives. Table 1 lists all probes and the expected behaviors from an OpenVPN server. An evaluation process is shown in Appendix Figure 11. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

FINGERPRINT

fingerprint.com

-- employees

Since n.d.

COMPANY RANKING

FINGERPRINT

EVERGREEN INDEX #1754

Representing this company? Click here to claim and customize this page!

Fingerprint WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!

FINGERPRINT

fingerprint.com

-- employees

Since n.d.

COMPANY RANKING

FINGERPRINT

EVERGREEN INDEX #1754

RELATED COMPANIESRANK

STORIES ABOUT FINGERPRINT

Representing this company? Click here to claim and customize this page!

Fingerprint WIKI

Apply to get your Tech Company News Page live on HackerNoon.com today!