Evaluation of Privacy for DNS Private Exchange

One of the IETF’s core views is that protocols should be designed to enable security and privacy while online . In light of the recent reported pervasive monitoring efforts, another goal is to design protocols and mechanisms to make such monitoring expensive or infeasible to conduct. As detailed in the DPRIVE problem statement , DNS resolution is an important arena for pervasive monitoring, and in some cases may be used for breaching the privacy of individuals. The set of DNS requests that an individual makes can provide a large amount of information about that individual. Not only individual requesters reveal information with their sets of DNS queries. In some specific use cases, the sets of DNS requests from a DNS recursive resolver or other entity may also provide revealing information. This document describes methods for measuring the performance of DNS privacy mechanisms; in particular, it provides methods for measuring effectiveness in the face of pervasive monitoring as defined in . The document includes example evaluations for common use cases. The privacy risks associated with DNS monitoring are not new, however they were brought into a greater visibility by the issue described in . The DPRIVE working group was formed to respond and at this time has several DNS private exchange mechanisms in consideration, including , , , and . There is also related work in other working groups, including DNSOP: and (potentially) DANE . The recently published also has relevance to DNS private exchange. Each effort related to DNS privacy mechanisms asserts some privacy assurances and operational relevance. Metrics for these privacy assurances are needed and are in reach based on existing techniques from the general field of privacy engineering. Systematic evaluation of DNS privacy mechanisms will enhance the likely operational effectiveness of DNS private exchange. Evaluating an individual mechanism for DNS privacy could be accomplished on a one-off basis, presumably as Privacy Considerations within each specification, but this will not address as much variation of operational contexts nor will it cover using multiple mechanisms together (in composition). Section 2 of discussed both benefits and risks of using multiple mechanisms. Definitions required for evaluating the privacy of stand-alone and composed design are not limited to privacy notions, but also need to include the risk model and some information about relationships among the entities in a given system. A mechanism for providing privacy to withstand the power and capabilities of a passive pervasive monitor may not withstand a more powerful actor using active monitoring by plugging itself into the path of individuals’ DNS requests as a forwarder . Having some standard models, and understanding how applicable they are to various designs is a part of evaluating the privacy. Sections 2 and 3 present privacy terminology and some assumptions. Sections 4 and 5 cover the system model or setup and the risk models of interest. In Section 6, we review a list of DNS privacy mechanisms, including some which are not in scope of the DPRIVE working group. Section 7 tackles how to evaluate privacy mechanisms, in the form of templates and outcomes. Given a specific risk model, the guarantees with respect to privacy of an individual or an item of interest are quantified.

This section provides definitions to be used for privacy evaluation of DNS. is the verbatim source of most of the definitions. Text is added to apply them to the DNS case. We follow the in classifying the terms. We have added a new section of terms to include several important practical or conventional terms that were not included in such as PII. For the terms from , we include their definitions rather than simply referencing them as an aid to readability.

Attacker: An entity that works against one or more privacy protection goals. Unlike observers, attackers' behavior is unauthorized, in a way similar to that of an eavesdropper. Eavesdropper: A type of attacker that passively observes an initiator's communications without the initiator's knowledge or authorization. This may include a passive pervasive monitor, defined below. Enabler: A protocol entity that facilitates communication between an initiator and a recipient without being directly in the communications path. DNS examples of an enabler in this sense include a recursive resolver, a proxy, or a forwarder. Individual: A human being (or a group of them) Initiator: A protocol entity that initiates communications with a recipient. Intermediary: A protocol entity that sits between the initiator (stub resolver) and the recipient (recursive resolver or authority resolver) and is necessary for the initiator and recipient to communicate. Unlike an eavesdropper, an intermediary is an entity that is part of the communication architecture and therefore at least tacitly authorized. Observer: An entity that is able to observe and collect information from communications, potentially posing privacy risks, depending on the context. As defined in this document, initiators, recipients, intermediaries, and enablers can all be observers. Observers are distinguished from eavesdroppers by being at least tacitly authorized. We note that while the definition of an observer may include an initiator in the risk model, an initiator of a request is excluded in the context of this document, because it corresponds to the subject of interest being studied. Similar to the definition in , we note that an attacker is broader than an observer. While claim that an attack does not consider the motive of the actor, the given context of DNS implies a motive if the term attacker is used to characterize the risk.

We assume the following definitions related to data and analysis from : attacker, correlation, fingerprint, fingerprinting, item of interest (IOI), personal data, interaction, traffic analysis, undetectability, and unlinkability. We augment some of those definitions later in this document. from , we relax the definition of IOI to exclude "the fact that a communication interaction has taken place" as this does not suite the evaluated context of DNS.

We assume the following definitions related to identifiability from : anonymity, anonymity set, anonymous, attribute, identity provider, personal name, and relying party. The following definitions are modified for the context of this document from those defined in Identifiability: The extent to which an individual is identifiable. has the rest of the variations on this (Identifiable, Identification, Identified, Identifier, Identity, Identity Confidentiality) Personal Name: A natural name for an individual. Personal names are often not unique and often comprise given names in combination with a family name. An individual may have multiple personal names at any time and over a lifetime, including official names. From a technological perspective, it cannot always be determined whether a given reference to an individual is, or is based upon, the individual's personal name(s) (see Pseudonym). NOTE: The reason to import this definition is that some query names that cause privacy leakage do so by embedding personal names as identifiers of host or other equipment, e.g. AllisonMankinMac.example.com. Pseudonymity: See the formal definition in the next section in lieu of . NOTE: Identifiability Definitions in also include some material not included here because the distinctions are not major for DNS Private Exchange, such as real and official names, and variant forms of Pseudonymity in its informal definition.

Central to the presentation of this document is the definition of personally identifiable information (PII), as well as other definitions that supplement the definitions listed earlier or modify them for the context of this document. In this section, we outline such definitions we further notes on their indications. Personally Identifiable Information (PII): Information (attributes) that can be used as is, or along with other side information, to identify, locate, and/or contact a single individual or subject (c.f. item of interest). NOTE: the definition above indicates that PII can be used on its own or in context. In DNS privacy, the items without additional context include IP(v4 or v6) address, qname, qtype, timings of queries, etc. The additional context includes organization-level attributes, such as a network prefix that can be associated with an organization. The definition of PII is complementary to the definition of items of interest. Subject: This term is useful as a parallel term to Individual. When the privacy of a group or an organization is of interest, we can reference the group or organization as Subject rather than Individual. Often it is desirable to reference alternative identifiers known as pseudonyms. A pseudonym is a name assumed by an individual in some context, unrelated to the names or identifiers known by others in that context. Pseudonymity/Pseudonym: a relaxation of the definition of anonymity for usability. In particular, pseudonymity is an anonymity feature obtained by using a pseudonym, an identifier that is used for establishing a long relationship between two entities. As an example, in the DNS context, a randomly generated pseudonym might identify a set of query data with a shared context, such as geographic origin. Such pseudonymity enables another entity interested in breaching the privacy to link multiple queries on a long-term basis. Pseudonyms are assumed long-lived and their uniqueness may be a goal. There are many findings that indicate that pseudonymity is weaker than anonymity. Unlinkability: Formally, two items of interest are said to be unlinkable if the certainty of an actor concerning those items of interest is not affected by observing the system. This is, unlinkability implies that the a-posteriori probability computed a monitor that two items of interest are related is close enough to the a-priori probability computed by a monitor based on his knowledge. Two items of interest are said to be unlinkable if there is a small (beta, close to 0) probability that the monitor identifies them as associated, and they are linkable if there is a sufficiently large probability (referred to as alpha). Informally, given two items of interest (user attributes, DNS queries, users, etc.), unlinkability is defined as the inability of the monitor to sufficiently determine whether those items are related to one another. In the context of DNS, this refers typically but not only to a monitor relating queries to the same individual. Undetectability: a stronger definition of privacy, where an item of interest is said to be undetectable if the monitor is not sufficiently able to know or tell whether the item exists or not. Note that undetectability implies unlinkability. As explained below, a way of ensuring undetectability is to use encryption secure under known ciphertext attacks, or randomized encryption. Unobservability: a stronger definition of privacy that requires satisfying both undetectability and anonymity. Unobservability means that an item of interest is undetectable by any uninvolved individual, monitor or not. In theory, there are many ways of ensuring unobservability by fulfilling both requirements. For example, undetectability requires that no party uninvolved in the resolution of a DNS query shall know that query has existed or not. A mechanism to ensure this function is encryption secure under known ciphertext attacks, or randomized encryption for all other than stub, and pseudonyms for the stub resolver. An alternative mechanism to provide the anonymity property would be the use of mix networks for routing DNS queries.

The quantification of privacy is connected with the privacy goals. Is the desired privacy property unlinkability only, or is it undetectability. Is pseudonymity a sufficient property? Parameters and entire privacy mechanism choices are affected by the choice of privacy goals. While a binary measure of privacy is sometimes possible, that is, being able to say that the transaction is anonymous, in this document, we assume that the binary is not frequently obtainable, and therefore we focus on methods for continuous quantification. Both are relevant to DNS Private Exchange. Another way to state this is that the quantification could be exactly the probabilities 1 and 0, corresponding to the binary, but the methods prefer to provide continuous values instead. Here is an example of continuous quantification, related to identifiability of an individual or item of interest based on observing queries. For an individual A, and a set of observations by a monitor, Y = [y1, y2, ... yn], we define the privacy of A as the uncertainty of the monitor of knowing that A is itself among many others under the observations Y; that is, we define Privacy = 1 - P[A | Y] For an item of interest r associated with a user A, we similarly define the privacy of r as Privacy = 1 – P[r | Y].

A DNS client (a DNS stub resolver) may resolve a domain name or address into the corresponding DNS record by contacting the authoritative name server responsible for that domain name (or address) directly. However, to improve the operation of DNS resolution, and reduce the round trip time required for resolving an address, both caching and recursive resolution are implemented. Caching is implemented at an intermediary between the stub and the authoritative name server. In practice, many caching servers also implement the recursive logic of DNS resolution for finding the name server authoritative for a domain, and are thus named DNS recursive resolvers. Another type of entity, forwarders (or proxies) are intermediaries between the three named here. The system model for DNS privacy evaluation includes the four entities quickly sketched here: stub resolvers, recursive resolvers, authoritative name servers, and forwarders.

Stub resolver (S): a minimal resolver that does not support referral, and delegates recursive resolution to a recursive resolver. A stub resolver is a consumer of recursive resolutions. Per the terminology of , a stub resolver is an Initiator. Recursive resolver (R): a resolver that implements the recursive function of DNS resolution on behalf of a stub resolver. Per the terminology of , a recursive resolver is an Enabler. Authoritative resolver (A): is a server that is the origin of a DNS record. A recursive resolver queries the authoritative resolver to resolve a domain name or address. Per the terminology of , the authoritative name server is also an Enabler. Forwarder/proxy (P): between the stub resolver and the authoritative resolver there may be more than one DNS-involved entity. These are systems located between S and R (stub resolver and recursive), or between R and A (recursive and authoritative), which do not play a primary role in the DNS protocol. Per the terminology of , forwarders are Intermediaries.

Evaluating various privacy protection mechanisms in relation to monitors such as the pervasive monitors defined next requires understanding links in the System setup. We define the following links. In relation to these are the attack surface where a monitor (eavesdropper) collects sets of query information. Stub -> Recursive (S-R): a link between the stub resolver and a recursive resolver. At the time of writing, the scope of DPRIVE Working Group privacy mechanisms is supposed to be limited to S-R. Stub -> Proxy (S-P): a link between the stub resolver and a forwarder/ proxy. The intended function of this link may be difficult to analyze. Proxy -> Recursive (P-R): a link between a proxy and a recursive server. Recursive -> Authoritative (R-A): a link between a recursive and an authoritative name server. Although at the time of writing, R-A is not in the DPRIVE scope, we touch on it in evaluations. Rather than notating in system setup that an entity is compromised, this is covered in the monitor model in Section 6, which has system elements as parameters. In the System Setup, there is a possibility that S and R exist on a single machine. The concept of the Unlucky Few relates S and R in this case. A monitor can monitor R-A and find the query traffic of the initiator individual. The same concept applies in the case where a recursive is serving a relatively small number of individuals. The query traffic of a subject group or organization (c.f. Subject in the definitions) is obtained by the monitor who monitors this system's R-A. Because R-A is not in the DPRIVE scope, it is for future work to examine the Unlucky Few circumstance fully. The general system setup is that PII, the individual’s private identifying information, is not sent on R-A and is not seen by authoritative name server. There could be one or more proxies between the stub resolver and a recursive. From a functionality point of view they can all be consolidated into a single proxy without affecting the system view, however, the behavior of such proxies may affect the size and shape of the attack surface. However, we believe that an additional treatment is needed for this case and it is not included in the discussion. We also do not include in discussion proxies that exist along R-A, between a recursive and an authoritative name server. We do so in respect for the DPRIVE charter's scope at this time. According to recent work at , there may be multiple intermediaries with poorly defined behavior. The system setup here leaves out other realistic considerations for simplicity, such as the impact of shared caches in DNS entities.

The Definitions section defines observer, attack and monitor, but not a Risk Model, which is needed to actually evaluate privacy, so this is now defined. For consistency, we note that the only difference between an attacker and an obeserver is that an attacker is an unauthorized observer with all the capabilities it may has. However, we also stress that for the context of DNS privacy, the term attacker may implicitly assume an intent. To that end, active and passive observers are collectively referred to as actors. Risk Model: a well-defined set of capabilities indicating how much information an observer (or eavesdropper) has, and in what context, in order to reach a goal of breaching the privacy of an individual or subject with respect to a given privacy metric. In this document we focus on two risk models, namely a pervasive monitor and a malicious monitor.

This risk corresponds to the passive pervasive monitoring model described in . This model relies on monitoring capabilities to breach the privacy of individuals from the DNS traffic at scale without decimation. An actor causing this risk is capable of eavesdropping or observing traffic between two end points, including traffic between any of the pairs of the entities described in section 2.1. Per , this type of actor has abilities to eavesdrop pervasively on many links at once, which is a powerful form of attack. Type-1 monitor are passive. They do not modify traffic or insert traffic.

an actor with the same types of capabilities of monitoring links, which selects links in order to target specific individuals. A Type-2 monitor for instance might put into place intermediaries in order to obtain traffic on specific links. Note that we exclude the malicious monitoring from this document since, by definition, a malicious actor has an intent associated with his actions.

To evaluate the privacy provided by a given mechanism or mechanisms in a particular system model, we characterize the risk with a template with parameters from the system model in which the risk actor (eavesdropper or observer as monitors) is located. The general template is: Risk(Type, [Entities], [Links]). For example, the template Risk(Type-2, R, S-R) passed as a parameter in the evaluation of a privacy mechanism indicates a Type-2 monitor that controls a recursive and has the capability of eavesdropping on the link between the stub and recursive resolvers. Other risk templates include the appropriate parameterizations based on the above description of those monitors, including monitors that have the capabilities of monitoring multiple links and controlling multiple pieces of infrastructure.

Various mechanisms for enhancing privacy in networks are applicable to DNS private exchange. Some mechanisms common to privacy research include mixing networks, dummy traffic, and private information retrieval techniques. Applicable protocol mechanisms include encryption-based techniques - encrypting the channel carrying the queries using IPSEC , TLS or special-purpose encryption . includes special-purpose encryption and also depends on a trusted service broker. Mixing Networks: in this type of mechanism, the initiator uses a mixing network such as Tor to route the DNS queries to the intended DNS server entity. A monitor observing part of the system finds it difficult to determine which individual sends which queries, and will not be able to tell which individual has sent them (ideally, though it is known that attacks exist that allow correlation and privacy breaches against mixing networks). The privacy property is unlinkability of the queries; the probability that two queries coming from one exit node in the mixing network belong to the same individual is uniform among all the individuals using the network. Dummy Traffic: a simple mechanism in which the initiator of a DNS request will also generate k dummy queries and send the intended query along with those queries. As such, the adversary will not be able to tell which query is of interest to the initiator. For a given k, the probability that the adversary will be able to detect which query is interest to the initiator is equal to 1-1/(k+1). In that sense, and for the proper parameterization of the protocol, the monitor is bounded to the undetectability of the queries. Private Information Retrieval: a mechanism that allows a user s to retrieve a record r from a database DB on a server without allowing the server to learn r. A trivial solution to the problem requires that s downloads the entire DB and then perform the queries locally. While that provides privacy to the queries of the user, the solution is communication inefficient at the scale of the DNS. More sophisticated cryptographic solutions are multi-round, and thus reduce the communication overhead, but are still inefficient for the DNS. Query Minimization: a mechanism that allows the resolver to minimize the amount of information it sends on behalf of a stub resolver. A method of query minimization is specified in . Qname minimization deprives a Type-1 risk on R-A of information from correlating queries, unless the individuals have an Unfortunate Few problem. NOTE: queries on R-A generally do not include an identifier of the individual making the query, because the source address is that of R. With respect R or A themselves, they may have well established policies for respecting the sensitivity of queries they process, while still using summary analysis of those queries to improve security, stability or their business operation. Encrypted Channel Mechanisms: Using these mechanisms, an initiator has an encrypted channel with a corresponding enabler, so that the queries are not available to eavesdropping Pervasive Monitor risk. Examples include , , and . Depending on the characteristics of the channel, various privacy properties are ensured. For instance, undetectability of queries is ensured for encryption-based mechanisms once the encrypted channel is established. Unlinkability of the queries may depend on the type of crypto-suite; it is provided as long as randomized encryption is used. Composed (Multiple) Mechanisms: the use of multiple mechanisms is a likely scenario and results in varied privacy guarantees. Consider a hypothetical system in which mixing networks (for unlinkability) and randomized encryption (for undetectability) can both be applied, thus providing for unobservability, a stronger property than either of the two along. On the other hand, consider another hypothetical system in which mixing networks are used to reach a third party broker requiring sign-in and authorization. Depending on the risk type, this could mean that the mixing network unlinkability was cancelled out by the linkability due to entrusting the third party with identifying information in order to be authorized.

Now we turn our attention to the evaluation of privacy mechanisms in a standard form, given the risk models and system definitions, for some of the example mechanisms. An evaluation takes multiple parameters as input. The output of the evaluation template is based on the analysis of the individual algorithms, settings, and parameters passed to this evaluation mechanism. Here is the top level interface of the evaluation template: Eval(Privacy_Mechanism(param_1, param_2, …), System_Setting(param_1, param_2, …), Risk_Model(param_1, param_2,...) The output of the function is a privacy guarantee for the given settings, expressed through defined properties such as unlinkability and unobservability, for the specified system and risk model. 7.1 Dummy Traffic Example Eval(Dummy_Traffic (k=10, distribution=uniform), System_Setting([S, P, R, A], [S-P, P-R, R-A]), Risk_Model(Type-1A, S-R)). The dummy traffic mechanism is not presented as a practical mechanism, though there's no way to know if there are deployments of this type of mechanism. This example evaluation uses k=10 to indicate that for every one query initiated by an individual, ten queries that disguise the query of interest are selected uniformly at random from a pool of queries. In the parameters passed in the evaluation function, we indicate that the privacy assurances of interest concern the S-R link, with a Passive Pervasive Monitor (Type-1A) risk. Here is a template format for the example:

Undetectability is provided with 0.91 probability (though we know there are other weaknesses for dummy traffic) If the threat model is replaced with Type-2, so that responses to arbitrary requests can be injected, and tracked, the undetectability probability is decreased. 7.2 Mixing Network Example Here is an input for a mixing network privacy mechanism: Eval(mix (u=10, distribution=uniform), System_Setting(link=S-R), threat_Model(Type-1A)). This indicates that the monitor resides between the stub and resolver. While queries are not undetectable, two queries are not linkable to the same individual; the provided guarantee is unlinkability. For a given number of individuals in the mixing network, indicated by the parameter u, assuming that at any time, traffic from these individuals is uniformly random, the probability that one query is comes from a given individual is (1/10=0.1). The probability that two queries are issued by the same initiator is 0.1^2 = 0.01, which represents the linkability probability. The unlinkability probability is given as 1-0.01 = 0.99. Thus, (unlinkability, 0.99) < Eval(mix (u=10, distribution=uniform), System_Setting(link=S-R), Risk_Model(type-1)). We note that even if there is a Type-2 Risk in R, the same results hold. To sum up, the above example is represented in the following template:

7.3 Encrypted Channel (DNS-over-TLS) Example For one of the encryption-based mechanisms, DNS-over-TLS , we have the following template (TLS parameters are from ):

This template features an Active Monitor risk model (Type-2) in order to show how that the monitor might apply extra resources to an encrypted channel. Undetectability is an issue whether using upgrade-based TLS on port 53, or a port-based TLS on a dedicated port - both ports indicate the use of DNS. The source address of the individual is exposed in all cases. If this were a suitably parameterized use of , the monitor would not be certain that all the traffic from S-R was DNS, and undetectability would be higher. 7.4 Encrypted Channel (IPSec) Example In the following, we use the same template above to characterize the encryption capabilities provided by IPSec, as a potential mechanisms for enabling privacy in DNS exchange.

We note that IPSec can provide better guarantees with respect to studied privacy notions. However, whether the technique itself is widely deployable or not is worth further investigation. 7.5 QName Minimization Example (R-A) Example Analyzing the privacy assurances of QName minimization is a non-trivial problem, given that the notions introduced in this document are techniques that do not alter items of interest. This is, the notions of privacy as outlined above are concerned with a certain IOI that is modified by this technique. To this end, we modify the aforementioned notions to suite this technique for analysis purpose only. For example, we define linkability as the ability of an adversary to link two labels of (minimized) queries to each other, and relate them to original source of query. Assuming a reasonable use of a recursive that minimizes queries on behalf of users, this task is non-trivial, although quantifying the probability would depend on the number of labels in queries, the number of queries issued, and the number of users using the studied recursive. The following template captures QName minimization as a template

Note that QName minimization does not solve the problem of the privacy for a monitoring risk between the stub and recursive. Encrypting the channel between the recursive and the stub, utilizing other techniques such as TDNS or IPSec, can marginalize such risk. Furthermore, note that the risk on the link between the recursive and authority name servers is always mitigated by the fact that recursive name servers act as a mixer of queries, even when they are sent in full to the authority name servers. 7.7 Private-DNS (S-R) Example The template for takes note of deployments in which in addition to S, R and A, there is another entity in the system, the function that authenticates the individual using S prior to permitting an encrypted channel to be formed to R or A. If the Private-DNS connection is with R, then identifiability of S as an individual may be similar to the identifiability of S from source address, or it may be stronger, depending on the nature of the account information required. If the Private-DNS connection is with A, source address PII is provided to A, and linkability of the queries from S has probability 1.

This document does not address a lot of the evaluation aspects not associated with privacy. For example, some of the mechanisms discussed in the working group are built of well-understood and standardized technologies, whereas others use other non-standard and less widely deployed techniques. A comprehensive evaluation of such mechanisms should take into account such facts.

The purpose of this document is to provide methods for those deploying or using DNS private exchange to assess the effectiveness of privacy mechanisms in depriving monitors of access to private information. Protecting privacy is one of the dimensions of an overall security strategy. It is possible for privacy-enhancing mechanisms to be deployed in ways that are vulnerable to security risks, with the result of not achieving security gains. For the purposes of privacy evaluation, it is important for the person making an evaluation to also ensure close attention to the content of the Security Considerations section of each mechanism being evaluated, for instance, to ensure if TLS is used for encryption of a link against surveillance, that TLS best security practices are in use.

No requests are made to IANA.

We wish to thank Scott Hollenbeck, Burt Kaliski, Minsuk Kang, Paul Livesay and Eric Osterweil for reviewing early versions. We wish to thank Stephane Bortzmeyer for his detailed review and feedback on the previous version of this document. We also wish to thank those who commented on presentations of this work ahead of publication, including Simson Garfinkel, Cathy Meadows, Paul Syverson, and Christine Task.