title: "Confidentiality in the Face of Pervasive Surveillance: A Threat Model and Problem Statement" abbrev: Confidentiality Threat Model docname: draft-iab-privsec-confidentiality-threat-00 date: 2014-09-11 category: info ipr: trust200902
ins: R. Barnes
name: Richard Barnes
email: [email protected]
- ins: B. Schneier name: Bruce Schneier email: [email protected]
- ins: C. Jennings name: Cullen Jennings email: [email protected]
- ins: T. Hardie name: Ted Hardie email: [email protected]
- ins: B. Trammell name: Brian Trammell email: [email protected]
- ins: C. Huitema name: Christian Huitema email: [email protected]
- ins: D. Borkmann name: Daniel Borkmann email: [email protected]
normative: RFC6973:
informative:
pass1:
target: http://www.theguardian.com/world/2013/jun/27/nsa-online-metadata-collection
title: "How the NSA is still harvesting your online data"
author:
organization: The Guardian
date: 2013
pass2:
target: http://www.theguardian.com/world/2013/jun/08/nsa-prism-server-collection-facebook-google
title: "NSA's Prism surveillance program: how it works and what it can do"
author:
organization: The Guardian
date: 2013
pass3:
target: http://www.theguardian.com/world/2013/jul/31/nsa-top-secret-program-online-data
title: "XKeyscore: NSA tool collects 'nearly everything a user does on the internet'"
author:
organization: The Guardian
date: 2013
pass4:
target: http://www.theguardian.com/uk/2013/jun/21/how-does-gchq-internet-surveillance-work
title: "How does GCHQ's internet surveillance work?"
author:
organization: The Guardian
dec1:
target: http://www.nytimes.com/2013/09/06/us/nsa-foils-much-internet-encryption.html
title: "N.S.A. Able to Foil Basic Safeguards of Privacy on Web"
author:
organization: The New York Times
date: 2013
dec2:
target: http://www.theguardian.com/world/interactive/2013/sep/05/nsa-project-bullrun-classification-guide
title: "Project Bullrun – classification guide to the NSA's decryption program"
author:
organization: The Guardian
date: 2013
dec3:
target: http://www.theguardian.com/world/2013/sep/05/nsa-gchq-encryption-codes-security
title: "Revealed: how US and UK spy agencies defeat internet privacy and security"
author:
organization: The Guardian
date: 2013
TOR:
target: https://www.torproject.org/
title: "Tor"
author:
organization: The Tor Project
date: 2013
TOR1:
target: https://www.schneier.com/blog/archives/2013/10/how_the_nsa_att.html
title: "How the NSA Attacks Tor/Firefox Users With QUANTUM and FOXACID"
author:
name: Bruce Schneier
ins: B. Schneier
date: 2013
TOR2:
target: http://www.theguardian.com/world/interactive/2013/oct/04/tor-stinks-nsa-presentation-document
title: "'Tor Stinks' presentation – read the full document"
author:
organization: The Guardian
date: 2013
dir1:
target: http://www.theguardian.com/world/2013/jun/06/nsa-phone-records-verizon-court-order
title: "NSA collecting phone records of millions of Verizon customers daily"
author:
organization: The Guardian
date: 2013
dir2:
target: http://www.theguardian.com/world/2013/jun/06/us-tech-giants-nsa-data
title: "NSA Prism program taps in to user data of Apple, Google and others"
author:
organization: The Guardian
date: 2013
dir3:
target: http://www.theguardian.com/world/interactive/2013/sep/05/sigint-nsa-collaborates-technology-companies
title: "Sigint – how the NSA collaborates with technology companies"
author:
organization: The Guardian
date: 2013
secure:
target: http://www.theguardian.com/world/2013/sep/05/nsa-how-to-remain-secure-surveillance
title: "NSA surveillance: A guide to staying secure"
author:
name: Bruce Schneier
ins: B. Schneier
organization: The Guardian
date: 2013
snowden:
target: http://www.technologyreview.com/news/519171/nsa-leak-leaves-crypto-math-intact-but-highlights-known-workarounds/
title: NSA Leak Leaves Crypto-Math Intact but Highlights Known Workarounds
author:
organization: Technology Review
date: 2013
key-recovery:
target: http://crypto.stanford.edu/~pgolle/papers/escrow.pdf
title: The Design and Implementation of Protocol-Based Hidden Key Recovery
author:
ins: E.-J. Goh
name: Eu-Jin Goh
author:
ins: D. Boneh
name: Dan Boneh
author:
ins: B. Pinkas
name: Benny Pinkas
author:
ins: P. Golle
name: Phillippe Golle
date: 2003
RFC1035:
RFC1918:
RFC1939:
RFC2015:
RFC2821:
RFC3261:
RFC3365:
RFC3501:
RFC3851:
RFC4033:
RFC4301:
RFC4303:
RFC4306:
RFC4949:
RFC5246:
RFC5321:
RFC5655:
RFC5750:
RFC6120:
RFC6962:
RFC6698:
RFC7011:
RFC7258:
--- abstract
Documents published in 2013 have revealed several classes of "pervasive" attack on Internet communications. In this document we develop a threat model that describes these pervasive attacks. We start by assuming a completely passive adversary with an interest in indiscriminate eavesdropping that can observe network traffic, then expand the threat model with a set of verified attacks that have been published. Based on this threat model, we discuss the techniques that can be employed in Internet protocol design to increase the protocols robustness to pervasive attacks.
--- middle
Starting in June 2013, documents released to the press by Edward Snowden have revealed several operations undertaken by intelligence agencies to exploit Internet communications for intelligence purposes. These attacks were largely based on protocol vulnerabilities that were already known to exist. The attacks were nonetheless striking in their pervasive nature, both in terms of the amount of Internet communications targeted, and in terms of the diversity of attack techniques employed.
To ensure that the Internet can be trusted by users, it is necessary for the Internet technical community to address the vulnerabilities exploited in these attacks {{RFC7258}}. The goal of this document is to describe more precisely the threats posed by these pervasive attacks, and based on those threats, lay out the problems that need to be solved in order to secure the Internet in the face of those threats.
The remainder of this document is structured as follows. In {{adversary}}, we describe an idealized passive adversary, one which could completely undetectably compromise communications at Internet scale. In {{reported}}, we provide a brief summary of some attacks that have been disclosed, and use these to expand the assumed capabilities of our idealized adversary. {{model}} describes a threat model based on these attacks, focusing on classes of attack that have not been a focus of Internet engineering to date. {{response}} provides some high-level guidance on how Internet protocols can defend against the threats described here.
This document makes extensive use of standard security and privacy terminology; see {{RFC4949}} and {{RFC6973}}. In addition, we use a few terms that are specific to the attacks discussed here:
Pervasive Attack: : An attack on Internet protocols that makes use of access at a large number of points in the network, or otherwise provides the attacker with access to a large amount of Internet traffic.
Observation: : Information collected directly from communications by an eavesdropper or observer. For example, the knowledge that <[email protected]> sent a message to <[email protected]> via SMTP taken from the headers of an observed SMTP message would be an observation.
Inference: :Information extracted from analysis of information collected directly from communications by an eavesdropper or observer. For example, the knowledge that a given web page was accessed by a given IP address, by comparing the size in octets of measured network flow records to fingerprints derived from known sizes of linked resources on the web servers involved, would be an inference.
Collaborator: : An entity that is a legitimate participant in a protocol, but who provides information about that interaction (keys or data) to an attacker.
Key Exfiltration: : The transmission of keying material for an encrypted communication from a collaborator to an attacker
Content Exfiltration: : The transmission of the content of a communication from a collaborator to an attacker
Unwitting Collaborator: : A collaborator that provides information to the attacker not deliberately, but because the attacker has exploited some technology used by the collaborator.
We assume a pervasive passive adversary, an indiscriminate eavesdropper on an Internet-attached computer network that
- can observe every packet of all communications at any or every hop in any network path between an initiator and a recipient; and
- can observe data at rest in intermediate systems between the endpoints controlled by the initiator and recipient; but
- takes no other action with respect to these communications (i.e., blocking, modification, injection, etc.).
This adversary is less capable than those which we know to have compromised the Internet from press reports, elaborated in {{reported}}, but represents the threat to communications privacy by a single entity interested in remaining undetectable.
The techniques available to our ideal adversary are direct observation and inference. Direct observation involves taking information directly from eavesdropped communications - e.g., URLs identifying content or email addresses identifying individuals from application-layer headers. Inference, on the other hand, involves analyzing eavesdropped information to derive new information from it; e.g., searching for application or behavioral fingerprints in observed traffic to derive information about the observed individual from them, in absence of directly-observed sources of the same information. The use of encryption to protect confidentiality is generally enough to prevent direct observation, assuming uncompromised encryption implementations and key material, but provides less complete protection against inference, especially inference based only on unprotected portions of communications (e.g. IP and TCP headers for TLS).
Protocols which do not encrypt their payload make the entire content of the communication available to a PPA along their path. Following the advice in {{RFC3365}}, most such protocols have a secure variant which encrypts payload for confidentiality, and these secure variants are seeing ever-wider deployment. A noteworthy exception is DNS {{RFC1035}}, as DNSSEC {{RFC4033}} does not have confidentiality as a requirement. This implies that all DNS queries and answers generated by the activities of any protocol are available to a the adversary.
Protocols which imply the storage of some data at rest in intermediaries leave this data subject to observation by an adversary that has compromised these intermediaries, unless the data is encrypted end-to-end by the application layer protocol, or the implementation uses an encrypted store for this data.
Inference is information extracted from later analysis of an observed communication, and/or correlation of observed information with information available from other sources. Indeed, most useful inference performed by a our ideal adversary falls under the rubric of correlation. The simplest example of this is the observation of DNS queries and answers from and to a source and correlating those with IP addresses with which that source communicates. This can give access to information otherwise not available from encrypted application payloads (e.g., the Host: HTTP/1.1 request header when HTTP is used with TLS).
Protocols which encrypt their payload using an application- or transport-layer encryption scheme (e.g. TLS {{RFC5246}}) still expose all the information in their network and transport layer headers to a PPA, including source and destination addresses and ports. IPsec ESP{{RFC4303}} further encrypts the transport-layer headers, but still leaves IP address information unencrypted; in tunnel mode, these addresses correspond to the tunnel endpoints. Features of the cryptographic protocols themselves, e.g. the TLS session identifier, may leak information that can be used for correlation and inference. While this information is much less semantically rich than the application payload, it can still be useful for the inferring an individual's activities.
Inference can also leverage information obtained from sources other than direct traffic observation. Geolocation databases, for example, have been developed map IP addresses to a location, in order to provide location-aware services such as targeted advertising. This location information is often of sufficient resolution that it can be used to draw further inferences toward identifying or profiling an individual.
Social media provide another source of more or less publicly accessible information. This information can be extremely semantically rich, including information about an individual's location, associations with other individuals and groups, and activities. Further, this information is generally contributed and curated voluntarily by the individuals themselves: it represents information which the individuals are not necessarily interested in protecting for privascy reasons. However, correlation of this social networking data with information available from direct observation of network traffic allows the creation of a much richer picture of an individual's activities than either alone. We note with some alarm that there is little that can be done from the protocol design side to limit such correlation by a PPA, and that the existence of such data sources in many cases greatly complicates the problem of protecting privacy by hardening protocols alone.
To illustrate how capable even this limited adversary is, we explore the non-anonymity of even encrypted IP traffic by examining in detail some inference techniques for associating a set of addresses with an individual, in order to illustrate the difficulty of defending communications against a PPA. Here, the basic problem is that information radiated even from protocols which have no obvious connection with personal data can be correlated with other information which can paint a very rich behavioral picture, that only takes one unprotected link in the chain to associate with an identity.
Internet traffic can be monitored by tapping Internet links, or by installing monitoring tools in Internet routers. Of course, a single link or a single router only provides access to a fraction of the global Internet traffic. However, monitoring a number of high capacity links or a set of routers placed at strategic locations provides access to a good sampling of Internet traffic.
Tools like IPFIX {{RFC7011}} allow administrators to acquire statistics about sequences of packets with some common properties that pass through a network device. The most common set of properties used in flow measurement is the "five-tuple"of source and destination addresses, protocol type, and source and destination ports. These statistics are commonly used for network engineering, but could certainly be used for other purposes.
Let's assume for a moment that IP addresses can be correlated to specific services or specific users. Analysis of the sequences of packets will quickly reveal which users use what services, and also which users engage in peer-to-peer connections with other users. Analysis of traffic variations over time can be used to detect increased activity by particular users, or in the case of peer-to-peer connections increased activity within groups of users.
The correlation of IP addresses with specific users can be done in various ways. For example, tools like reverse DNS lookup can be used to retrieve the DNS names of servers. Since the addresses of servers tend to be quite stable and since servers are relatively less numerous than users, a PPA could easily maintain its own copy of the DNS for well-known or popular servers, to accelerate such lookups.
On the other hand, the reverse lookup of IP addresses of users is generally less informative. For example, a lookup of the address currently used by one author's home network returns a name of the form "c-192-000-002-033.hsd1.wa.comcast.net". This particular type of reverse DNS lookup generally reveals only coarse-grained location or provider information.
In many jurisdictions, Internet Service Providers (ISPs) are required to provide identification on a case by case basis of the "owner" of a specific IP address for law enforcement purposes. This is a reasonably expedient process for targeted investigations, but pervasive surveillance requires something more efficient. This provides an incentive for the adversary to secure the cooperation of the ISP in order to automate this correlation.
Even if the ISP does not cooperate, user identity can often be obtained via inference. POP3 {{RFC1939}} and IMAP {{RFC3501}} are used to retrieve mail from mail servers, while a variant of SMTP {{RFC5321}} is used to submit messages through mail servers. IMAP connections originate from the client, and typically start with an authentication exchange in which the client proves its identity by answering a password challenge. The same holds for the SIP protocol {{RFC3261}} and many instant messaging services operating over the Internet using proprietary protocols.
The username is directly observable if any of these protocols operate in cleartext; the username can then be directly associated with the source address.
SMTP {{RFC5321}} requires that each successive SMTP relay adds a "Received" header to the mail headers. The purpose of these headers is to enable audit of mail transmission, and perhaps to distinguish between regular mail and spam. Here is an extract from the headers of a message recently received from the "perpass" mailing list:
Received: from 192-000-002-044.zone13.example.org (HELO ?192.168.1.100?)
(xxx.xxx.xxx.xxx) by lvps192-000-002-219.example.net with ESMTPSA
(DHE-RSA-AES256-SHA encrypted, authenticated); 27 Oct 2013 21:47:14 +0100
Message-ID: <[email protected]>
Date: Sun, 27 Oct 2013 20:47:14 +0000
From: Some One <[email protected]>
This is the first "Received" header attached to the message by the first SMTP relay; for privacy reasons, the field values have been anonymized. We learn here that the message was submitted by "Some One" on October 27, from a host behind a NAT (192.168.1.100) {{RFC1918}} that used the IP address 192.0.2.44. The information remained in the message, and is accessible by all recipients of the "perpass" mailing list, or indeed by any PPA that sees at least one copy of the message.
An idealized adversary that can observe sufficient email traffic can regularly update the mapping between public IP addresses and individual email identities. Even if the SMTP traffic was encrypted on submission and relaying, the adversary can still receive a copy of public mailing lists like "perpass".
Many web sites only encrypt a small fraction of their transactions. A popular pattern was to use HTTPS for the login information, and then use a "cookie" to associate following clear-text transactions with the user's identity. Cookies are also used by various advertisement services to quickly identify the users and serve them with "personalized" advertisements. Such cookies are particularly useful if the advertisement services want to keep tracking the user across multiple sessions that may use different IP addresses.
As cookies are sent in clear text, a PPA can build a database that associates cookies to IP addresses for non-HTTPS traffic. If the IP address is already identified, the cookie can be linked to the user identify. After that, if the same cookie appears on a new IP address, the new IP address can be immediately associated with the pre-determined identity.
An adversary can track traffic from an IP address not yet associated with an individual to various public services (e.g. websites, mail servers, game servers), and exploit patterns in the observed traffic to correlate this address with other addresses that show similar patterns. For example, any two addresses that show connections to the same IMAP or webmail services, the same set of favorite websites, and game servers at similar times of day may be associated with the same individual. Correlated addresses can then be tied to an individual through one of the techniques above, walking the "network graph" to expand the set of attributable traffic.
The situation in reality is more bleak than that suggested by an analysis of our idealized adversary. Through revelations of sensitive documents in several media outlets, the Internet community has been made aware of several intelligence activities conducted by US and UK national intelligence agencies, particularly the US National Security Agency (NSA) and the UK Government Communications Headquarters (GCHQ). These documents have revealed methods that these agencies use to attack Internet applications and obtain sensitive user information.
First, they have confirmed that these agencies have capabilities in line with those of our idealized adversary, thorugh the large-scale passive collection of Internet traffic {{pass1}}{{pass2}}{{pass3}}{{pass4}}. For example:
* The NSA XKEYSCORE system accesses data from multiple access points and searches for "selectors" such as email addresses, at the scale of tens of terabytes of data per day.
* The GCHQ Tempora system appears to have access to around 1,500 major cables passing through the UK.
* The NSA MUSCULAR program tapped cables between data centers belonging to major service providers.
* Several programs appear to perform wide-scale collection of cookies in web traffic and location data from location-aware portable devices such as smartphones.
However, the capabilities described go beyond those available to our idealized adversary, including:
-
Decryption of TLS-protected Internet sessions {{dec1}}{{dec2}}{{dec3}}. For example, the NSA BULLRUN project appears to have had a budget of around $250M per year to undermine encryption through multiple approaches.
-
Insertion of NSA devices as a man-in-the-middle of Internet transactions {{TOR1}}{{TOR2}}. For example, the NSA QUANTUM system appears to use several different techniques to hijack HTTP connections, ranging from DNS response injection to HTTP 302 redirects.
-
Direct acquisition of bulk data and metadata from service providers {{dir1}}{{dir2}}{{dir3}}. For example, the NSA PRISM program provides the agency with access to many types of user data (e.g., email, chat, VoIP).
-
Use of implants (covert modifications or malware) to undermine security and anonymity features {{dec2}}{{TOR1}}{{TOR2}}. For example:
- NSA appears to use the QUANTUM man-in-the-middle system to direct users to a FOXACID server, which delivers an implant to compromise the browser of a user of the Tor anonymous communications network.
- The BULLRUN program mentioned above includes the addition of covert modifications to software as one means to undermine encryption.
- There is also some suspicion that NSA modifications to the DUAL_EC_DRBG random number generator were made to ensure that keys generated using that generator could be predicted by NSA. These suspicions have been reinforced by reports that RSA Security was paid roughly $10M to make DUAL_EC_DRBG the default in their products.
We use the term "pervasive attack" to collectively describe these operations. The term "pervasive" is used because the attacks are designed to indiscriminately gather as much data as possible and to apply selective analysis on targets after the fact. This means that all, or nearly all, Internet communications are targets for these attacks. To achieve this scale, the attacks are physically pervasive; they affect a large number of Internet communications. They are pervasive in content, consuming and exploiting any information revealed by the protocol. And they are pervasive in technology, exploiting many different vulnerabilities in many different protocols.
It's important to note that although the attacks mentioned above were executed by NSA and GCHQ, there are many other organizations that can mount pervasive attacks. Because of the resources required to achieve pervasive scale, pervasive attacks are most commonly undertaken by nation-state actors. For example, the Chinese Internet filtering system known as the "Great Firewall of China" uses several techniques that are similar to the QUANTUM program, and which have a high degree of pervasiveness with regard to the Internet in China.
Given these disclosures, we must consider broader threat model.
Pervasive surveillance aims to collect information across a large number of Internet communications, observing the collected communications to identify information of interest within individual communications, or inferring information from correlated communications. This analysis sometimes benefits from decryption of encrypted communications and deanonymization of anonymized communications. As a result, these attackers desire both access to the bulk of Internet traffic and to the keying material required to decrypt any traffic that has been encrypted (though the presence of a communication and the fact that it is encrypted may both be inputs to an analysis, even if the attacker cannot decrypt the communication).
The attacks listed above highlight new avenues both for access to traffic and for access to relevant encryption keys. They further indicate that the scale of surveillance is sufficient to provide a general capability to cross-correlate communications, a threat not previously thought to be relevant at the scale of all Internet communications.
| Attack Class | Capability |
|---|---|
| Passive observation | Directly capture data in transit |
| Passive inference | Infer from reduced/encrypted data |
| Active | Manipulate / inject data in transit |
| Static key exfiltration | Obtain key material once / rarely |
| Dynamic key exfiltration | Obtain per-session key material |
| Content exfiltration | Access data at rest |
Security analyses of Internet protocols commonly consider two classes of attacker: Passive attackers, who can simply listen in on communications as they transit the network, and "active attackers", who can modify or delete packets in addition to simply collecting them.
In the context of pervasive attack, these attacks take on an even greater significance. In the past, these attackers were often assumed to operate near the edge of the network, where attacks can be simpler. For example, in some LANs, it is simple for any node to engage in passive listening to other nodes' traffic or inject packets to accomplish active attacks. In the pervasive attack case, however, both passive and active attacks are undertaken closer to the core of the network, greatly expanding the scope and capability of the attacker.
A passive attacker with access to a large portion of the Internet can analyze collected traffic to create a much more detailed view of user behavior than an attacker that collects at a single point. Even the usual claim that encryption defeats passive attackers is weakened, since a pervasive passive attacker can infer relationships from correlations over large numbers of sessions, e.g., pairing encrypted sessions with unencrypted sessions from the same host, or performing traffic fingerprinting between known and unknown encrypted sessions. The reports on the NSA XKEYSCORE system would make it an example of such an attacker.
A pervasive active attacker likewise has capabilities beyond those of a localized active attacker. Active attacks are often limited by network topology, for example by a requirement that the attacker be able to see a targeted session as well as inject packets into it. A pervasive active attacker with multiple accesses at core points of the Internet is able to overcome these topological limitations and apply attacks over a much broader scope. Being positioned in the core of the network rather than the edge can also enable a pervasive active attacker to reroute targeted traffic. Pervasive active attackers can also benefit from pervasive passive collection to identify vulnerable hosts.
While not directly related to pervasiveness, attackers that are in a position to mount a pervasive active attack are also often in a position to subvert authentication, the traditional response to active attack. Authentication in the Internet is often achieved via trusted third party authorities such as the Certificate Authorities (CAs) that provide web sites with authentication credentials. An attacker with sufficient resources for pervasive attack may also be able to induce an authority to grant credentials for an identity of the attacker’s choosing. If the parties to a communication will trust multiple authorities to certify a specific identity, this attack may be mounted by suborning any one of the authorities (the proverbial "weakest link"). Subversion of authorities in this way can allow an active attack to succeed in spite of an authentication check.
Beyond these three classes (observation, inference, and active), reports on the BULLRUN effort to defeat encryption and the PRISM effort to obtain data from service providers suggest three more classes of attack:
- Static key exfiltration
- Dynamic key exfiltration
- Content exfiltration
These attacks all rely on a "collaborator" endpoint providing the attacker with some information, either keys or data. These attacks have not traditionally been considered in security analyses of protocols, since they happen outside of the protocol.
The term "key exfiltration" refers to the transfer of keying material for an encrypted communication from the collaborator to the attacker. By "static", we mean that the transfer of keys happens once, or rarely, typically of a long-lived key. For example, this case would cover a web site operator that provides the private key corresponding to its HTTPS certificate to an intelligence agency.
"Dynamic" key exfiltration, by contrast, refers to attacks in which the collaborator delivers keying material to the attacker frequently, e.g., on a per-session basis. This does not necessarily imply frequent communications with the attacker; the transfer of keying material may be virtual. For example, if an endpoint were modified in such a way that the attacker could predict the state of its psuedorandom number generator, then the attacker would be able to derive per-session keys even without per-session communications.
Finally, content exfiltration is the attack in which the collaborator simply provides the attacker with the desired data or metadata. Unlike the key exfiltration cases, this attack does not require the attacker to capture the desired data as it flows through the network. The risk is to data at rest as opposed to data in transit. This increases the scope of data that the attacker can obtain, since the attacker can access historical data -- the attacker does not have to be listening at the time the communication happens.
Exfiltration attacks can be accomplished via attacks against one of the parties to a communication, i.e., by the attacker stealing the keys or content rather than the party providing them willingly. In these cases, the party may not be aware that they are collaborating, at least at a human level. Rather, the subverted technical assets are "collaborating" with the attacker (by providing keys/content) without their owner's knowledge or consent.
Any party that has access to encryption keys or unencrypted data can be a collaborator. While collaborators are typically the endpoints of a communication (with encryption securing the links), intermediaries in an unencrypted communication can also facilitate content exfiltration attacks as collaborators by providing the attacker access to those communications. For example, documents describing the NSA PRISM program claim that NSA is able to access user data directly from servers, where it was stored unencrypted. In these cases, the operator of the server would be a collaborator (wittingly or unwittingly). By contrast, in the NSA MUSCULAR program, a set of collaborators enabled attackers to access the cables connecting data centers used by service providers such as Google and Yahoo. Because communications among these data centers were not encrypted, the collaboration by an intermediate entity allowed NSA to collect unencrypted user data.
| Attack Class | Cost / Risk to Attacker |
|---|---|
| Passive observation | Passive data access |
| Passive inference | Passive data access + processing |
| Active | Active data access + processing |
| Static key exfiltration | One-time interaction |
| Dynamic key exfiltration | Ongoing interaction / code change |
| Content exfiltration | Ongoing, bulk interaction |
In order to realize an attack of each of the types discussed above, the attacker has to incur certain costs and undertake certain risks. These costs differ by attack, and can be helpful in guiding response to pervasive attack.
Depending on the attack, the attacker may be exposed to several types of risk, ranging from simply losing access to arrest or prosecution. In order for any of these negative consequences to happen, however, the attacker must first be discovered and identified. So the primary risk we focus on here is the risk of discovery and attribution.
A passive attack is the simplest attack to mount in some ways. The base requirement is that the attacker obtain physical access to a communications medium and extract communications from it. For example, the attacker might tap a fiber-optic cable, acquire a mirror port on a switch, or listen to a wireless signal. The need for these taps to have physical access or proximity to a link exposes the attacker to the risk that the taps will be discovered. For example, a fiber tap or mirror port might be discovered by network operators noticing increased attenuation in the fiber or a change in switch configuration. Of course, passive attacks may be accomplished with the cooperation of the network operator, in which case there is a risk that the attacker's interactions with the network operator will be exposed.
In many ways, the costs and risks for an active attack are similar to those for a passive attack, with a few additions. An active attacker requires more robust network access than a passive attacker, since for example they will often need to transmit data as well as receiving it. In the wireless example above, the attacker would need to act as an transmitter as well as receiver, greatly increasing the probability the attacker will be discovered (e.g., using direction-finding technology). Active attacks are also much more observable at higher layers of the network. For example, an active attacker that attempts to use a mis-issued certificate could be detected via Certificate Transparency {{RFC6962}}.
In terms of raw implementation complexity, passive attacks require only enough processing to extract information from the network and store it. Active attacks, by contrast, often depend on winning race conditions to inject pakets into active connections. So active attacks in the core of the network require processing hardware to that can operate at line speed (roughly 100Gbps to 1Tbps in the core) to identify opportunities for attack and insert attack traffic in a high-volume traffic.
Key exfiltration attacks rely on passive attack for access to encrypted data, with the collaborator providing keys to decrypt the data. So the attacker undertakes the cost and risk of a passive attack, as well as additional risk of discovery via the interactions that the attacker has with the collaborator.
In this sense, static exfiltration has a lower risk profile than dynamic. In the static case, the attacker need only interact with the collaborator a small number of times, possibly only once, say to exchange a private key. In the dynamic case, the attacker must have continuing interactions with the collaborator. As noted above these interactions may real, such as in-person meetings, or virtual, such as software modifications that render keys available to the attacker. Both of these types of interactions introduce a risk that they will be discovered, e.g., by employees of the collaborator organization noticing suspicious meetings or suspicious code changes.
Content exfiltration has a similar risk profile to dynamic key exfiltration. In a content exfiltration attack, the attacker saves the cost and risk of conducting a passive attack. The risk of discovery through interactions with the collaborator, however, is still present, and may be higher. The content of a communication is obviously larger than the key used to encrypt it, often by several orders of magnitude. So in the content exfiltration case, the interactions between the collaborator and the attacker need to be much higher-bandwidth than in the key exfiltration cases, with a corresponding increase in the risk that this high-bandwidth channel will be discovered.
It should also be noted that in these latter three exfiltration cases, the collaborator also undertakes a risk that his collaboration with the attacker will be discovered. Thus the attacker may have to incur additional cost in order to convince the collaborator to participate in the attack. Likewise, the scope of these attacks is limited to case where the attacker can convince a collaborator to participate. If the attacker is a national government, for example, it may be able to compel participation within its borders, but have a much more difficult time recruiting foreign collaborators.
As noted above, the "collaborator" in an exfiltration attack can be unwitting; the attacker can steal keys or data to enable the attack. In some ways, the risks of this approach are similar to the case of an active collaborator. In the static case, the attacker needs to steal information from the collaborator once; in the dynamic case, the attacker needs to continued presence inside the collaborators systems. The main difference is that the risk in this case is of automated discovery (e.g., by intrusion detection systems) rather than discovery by humans.
Given this threat model, how should the Internet technical community respond to pervasive attack?
The cost and risk considerations discussed above can provide a guide to response. Namely, responses to passive attack should close off avenues for attack that are safe, scalable, and cheap, forcing the attacker to mount attacks that expose it to higher cost and risk.
In this section, we discuss a collection of high-level approaches to mitigating pervasive attacks. These approaches are not meant to be exhaustive, but rather to provide general guidance to protocol designers in creating protocols that are resistant to pervasive attack.
| Attack Class | High-level mitigations |
|---|---|
| Passive observation | Encryption for confidentiality |
| Passive inference | ??? |
| Active | Authentication, monitoring |
| Static key exfiltration | Encryption with per-session state (PFS) |
| Dynamic key exfiltration | Transparency, validation of end systems |
| Content exfiltration | Object encryption, distributed systems |
The traditional mitigation to passive attack is to render content unintelligible to the attacker by applying encryption, for example, by using TLS or IPsec {{RFC5246}}{{RFC4301}}. Even without authentication, encryption will prevent a passive attacker from being able to read the encrypted content. Exploiting unauthenticated encryption requires an active attack (man in the middle); with authentication, a key exfiltration attack is required.
The additional capabilities of a pervasive passive attacker, however, require some changes in how protocol designers evaluate what information is encrypted. In addition to directly collecting unencrypted data, a pervasive passive attacker can also make inferences about the content of encrypted messages based on what is observable. For example, if a user typically visits a particular set of web sites, then a pervasive passive attacker observing all of the user's behavior can track the user based on the hosts the user communicates with, even if the user changes IP addresses, and even if all of the connections are encrypted.
Thus, in designing protocols to be resistant to pervasive passive attacks, protocol designers should consider what information is left unencrypted in the protocol, and how that information might be correlated with other traffic. Information that cannot be encrypted should be anonymized, i.e., it should be dissociated from other information. For example, the Tor overlay routing network anonymizes IP addresses by using multi-hop onion routing {{TOR}}.
As with traditional, limited active attacks, the basic mitigation to pervasive active attack is to enable the endpoints of a communication to authenticate each other. However, as noted above, attackers that can mount pervasive active attacks can often subvert the authorities on which authentication systems rely. Thus, in order to make authentication systems more resilient to pervasive attack, it is beneficial to monitor these authorities to detect misbehavior that could enable active attack. For example, DANE and Certificate Transparency both provide mechanisms for detecting when a CA has issued a certificate for a domain name without the authorization of the holder of that domain name {{RFC6962}}{{RFC6698}}.
While encryption and authentication protect the security of individual sessions, these sessions may still leak information, such as IP addresses or server names, that a pervasive attacker can use to correlate sessions and derive additional information about the target. Thus, pervasive attack highlights the need for anonymization technologies, which make correlation more difficult. Typical approaches to anonymization against traffic analysis include:
- Aggregation: Routing sessions for many endpoints through a common mid-point (e.g., an HTTP proxy). Since the midpoint appears as the end of the communication, individual endpoints cannot be distinguished.
- Onion routing: Routing a session through several mid-points, rather than directly end-to-end, with encryption that guarantees that each node can only see the previous and next hops {{TOR}}. This ensures that the source and destination of a communication are never revealed simultaneously.
- Multi-path: Routing different sessions via different paths (even if they originate from the same endpoint). This reduces the probability that the same attacker will be able to collect many sessions.
An encrypted, authenticated session is safe from content-monitoring attacks in which neither end collaborates with the attacker, but can still be subverted by the endpoints. The most common ciphersuites used for HTTPS today, for example, are based on using RSA encryption in such a way that if an attacker has the private key, the attacker can derive the session keys from passive observation of a session. These ciphersuites are thus vulnerable to a static key exfiltration attack -- if the attacker obtains the server's private key once, then they can decrypt all past and future sessions for that server.
Static key exfiltration attacks are prevented by including ephemeral, per-session secret information in the keys used for a session. Most IETF security protocols include modes of operation that have this property. These modes are known in the literature under the heading "perfect forward secrecy" (PFS) because even if an adversary has all of the secrets for one session, the next session will use new, different secrets and the attacker will not be able to decrypt it. The Internet Key Exchange (IKE) protocol used by IPsec supports PFS by default {{RFC4306}}, and TLS supports PFS via the use of specific ciphersuites {{RFC5246}}.
Dynamic key exfiltration cannot be prevent by protocol means. By definition, any secrets that are used in the protocol will be transmitted to the attacker and used to decrypt what the protocol encrypts. Likewise, no technical means will stop a willing collaborator from sharing keys with an attacker. However, this attack model also covers "unwitting collaborators", whose technical resources are collaborating with the attacker without their owners' knowledge. This could happen, for example, if flaws are built into products or if malware is injected later on.
The best defense against becoming an unwitting collaborator is thus to assure that end systems are well-vetted and secure. Transparency is a major tool in this process {{secure}}. Open source software is easier to evaluate for potential flaws than proprietary software, by a wider array of independent analysts. Products that conform to standards for cryptography and security protocols are limited in the ways they can misbehave. And standards processes that are open and transparent help ensure that the standards themselves do not provide avenues for attack.
Standards can also define protocols that provide greater or lesser opportunity for dynamic key exfiltration. Collaborators engaging in key exfiltration through a standard protocol will need to use covert channels in the protocol to leak information that can be used by the attacker to recover the key. Such use of covert channels has been demonstrated for SSL, TLS, and SSH {{key-recovery}}. Any protocol bits that can be freely set by the collaborator can be used as a covert channel, including, for example, TCP options or unencrypted traffic sent before a STARTTLS message in SMTP or XMPP. Protocol designers should consider what covert channels their protocols expose, and how those channels can be exploited to exfiltrate key information.
Content exfiltration has some similarity to the dynamic exfiltration case, in that nothing can prevent a collaborator from revealing what they know, and the mitigations against becoming an unwitting collaborator apply. In this case, however, applications can limit what the collaborator is able to reveal. For example, the S/MIME and PGP systems for secure email both deny intermediate servers access to certain parts of the message {{RFC5750}}{{RFC2015}}. Even if a server were to provide an attacker with full access, the attacker would still not be able to read the protected parts of the message.
Mechanisms like S/MIME and PGP are often referred to as "end-to-end" security mechanisms, as opposed to "hop-by-hop" or "end-to-middle" mechanisms like the use of SMTP over TLS. These two different mechanisms address different types of attackers: Hop-by-hop mechanisms protect from attackers on the wire (passive or active), while end-to-end mechansims protect against attackers within intermediate nodes. Thus, neither of these mechanisms provides complete protection by itself. For example:
-
Two users messaging via Facebook over HTTPS are protected against passive and active attackers in the network between the users and Facebook. However, if Facebook is a collaborator in an exfiltration attack, their communications can still be monitored. They would need to encrypt their messages end-to-end in order to protect themselves against this risk.
-
Two users exchanging PGP-protected email have protected the content of their exchange from network attackers and intermediate servers, but the header information (e.g., To and From addresses) is unnecessarily exposed to passive and active attackers that can see communications among the mail agents handling the email messages. These mail agents need to use hop-by-hop encryption and traffic analysis mitigation to address this risk.
Mechanisms such as S/MIME and PGP are also known as "object-based" security mechanisms (as opposed to "communications security" mechanisms), since they operate at the level of objects, rather than communications sessions. Such secure object can be safely handled by intermediaries in order to realize, for example, store and forward messaging. In the examples above, the encrypted instant messages or email messages would be the secure objects.
The mitigations to the content exfiltration case are thus to regard participants in the protocol as potential passive attackers themselves, and apply the mitigations discussed above with regard to passive attack. Information that is not necessary for these participants to fulfill their role in the protocol can be encrypted, and other information can be anonymized.
In summary, many of the basic tools for mitigating pervasive attack already exist. As Edward Snowden put it, "properly implemented strong crypto systems are one of the few things you can rely on" {{snowden}}. The task for the Internet community is to ensure that applications are able to use the strong crypto systems we have defined -- for example, TLS with PFS ciphersuites -- and that these properly implemented. (And, one might add, turned on!) Some of this work will require architectural changes to applications, e.g., in order to limit the information that is exposed to servers. In many other cases, however, the need is simply to make the best use we can of the cryptographic tools we have.
- Thaler for list of attacks and taxonomy
- Security ADs for starting and managing the perpass discussion
- See PPA acks as well
- Ensure all bases are covered WRT threats to confidentiality
- Consider moving mitigations to a separate document per program description
- Look at better alignment with draft-farrell-perpass-attack
- Better coverage of traffic analysis -- PPA helped somewhat here but the problem is hard
- Terminology alignment (after the program agrees the structure is good)