DNS privacy problem statement

The Domain Name System is specified in and . It is one of the most important infrastructure components of the Internet and one of the most often ignored or misunderstood. Almost every activity on the Internet starts with a DNS query (and often several). Its use has many privacy implications and we try to give here a comprehensive and correct list. Let us start with a small reminder of the way the DNS works (with some simplifications). A client, the stub resolver, issues a DNS query to a server, the resolver. For instance, the query is "What are the AAAA records for www.example.com?". AAAA is the qtype (Query Type) and www.example.com the qname (Query Name). To get the answer, the resolver will query first the root nameservers, which will, most of the times, send a referral. Here, the referral will be to .com nameservers. In turn, they will send a referral to the example.com nameservers, which will provide the answer. The root name servers, the name servers of .com and those of example.com are called authoritative name servers. It is important, when analyzing the privacy issues, to remember that the question asked to all these name servers is always the original question, not a derived question. Unlike what many "DNS for dummies" articles say, the question sent to the root name servers is "What are the AAAA records for www.example.com?", not "What are the name servers of .com?". So, the DNS leaks more information than it should. Because the DNS uses caching heavily, not all questions are sent to the authoritative name servers. If the stub resolver, a few seconds later, ask to the resolver "What are the SRV records of _xmpp-server._tcp.example.com?", the resolver will remember that it knows the name servers of example.com and will just query them, bypassing the root and .com. Because there is typically no caching in the stub resolver, the resolver, unlike the authoritative servers, sees everything. Almost all the DNS queries are today sent over UDP, which have practical consequences if someones thinks of encrypting this traffic. I should be noted to that DNS resolvers often forward requests to a bigger machine, with a larger and more shared cache, the forwarders. From the point of view of privacy, forwarders are like resolvers, except that the caching in the resolver before them decreases the amount of data they can see. We will use here the terminology of .

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in .

The DNS request includes many fields but two seem specially relevant for the privacy issues, the qname and the source IP address. The qname is the full question sent by the user. It gives information about what the user does ("What are the MX records of example.net?" means he probably wants to send email to someone at example.net, which may be a domain used by only a few persons and therefore very revealing). Some qnames are more sensitive than others. For instance, querying the A record of google-analytics.com reveals very little (everybody visit Web sites which use Google Analytics) but querying the A record of www.verybad.example where verybad.example is the domain of an illegal or very offensive organization may create more problems for the user. Another example if the qname embedding the software you use. For instance, some BitTorrent clients query a SRV record for _bittorrent-tracker._tcp.domain.example. For the communication between the stub resolver and the resolver, the source IP address is the one of the user's machine. Therefore, all the issues and warnings about collection of IP addresses apply here. For the communication between the resolver and the authoritative name servers, the source IP address has a different meaning, it does not have the same status as the source address in a HTTP connection. It is now the IP address of the resolver which, in a way "hides" the real user. However, it does not always work. Sometimes the end user has a personal resolver on her machine. In that case, the IP address is as sensitive as it is for HTTP.

DNS traffic can be seen by an eavesdropper like any other traffic. It is typically not encrypted. (DNSSEC, specified in explicitely excludes confidentiality from its goals.) So, if an initiator starts a HTTPS communication with a recipient, while the HTTP traffic will be encrypted, the DNS exchange prior to it won't be. What also makes the DNS traffic different is that it may take a different path than the communication between the initiator and the recipient. For instance, an eavesdropper may be unable to tap the wire between the initiator and the recipient but may have access to the wire going to the resolver, or to the authoritative name servers. The best place, from an eavesdropper's point of view, is clearly between the stub resolvers and the resolvers, because you are not limited by DNS caching.

Using the terminology of , the DNS servers (resolvers and authoritative servers) are enablers: they facilitate communication between an initiator and a recipient without being directly in the communications path. As a result, they are often forgotten in risk analysis. But, to quote again , "Although [...] enablers may not generally be considered as attackers, they may all pose privacy threats (depending on the context) because they are able to observe, collect, process, and transfer privacy-relevant data." In parlance, enablers become observers when they start collecting data. Many programs exist to collect and analyze DNS data at the servers. From the "query log" of some programs like BIND, to tcpdump and more sophisticated programs like PacketQ and DNSmezzo. The organization managing the DNS server can use this data itself or it can be part of a surveillance program like PRISM and pass data to an outside attacker. Sometimes, these data are kept for a long time and/or distributed to third parties, for research purposes [ditl], for security analysis, or for surveillance tasks.

The resolvers see the entire traffic since there is typically no caching before them. They are therefore well situated to observe the traffic. To summary: your resolver knows a lot about you. The resolver of a big ISP, or a big public resolver can collect data from many users. You may get an idea of the data collected by reading the privacy policy of a big public resolver.

Unlike the resolvers, they are limited by caching. They see only a part of the requests. For aggregated statistics ("what is the percentage of LOC queries?"), it is sufficient but it may prevent an observer to observe everything. Nevertheless, the authoritative name servers see a random part of the traffic and it may be sufficient to defeat some privacy expectations. Also, the end user has typically some legal/contractual link with the resolver (he choosed the ISP, or he choosed to use a given public resolver) while he is often not even aware of the role of the authoritative name servers and their observation abilities. It is an interesting question whether the privacy issues are bigger in the root or in a big TLD. The root sees the traffic for all the TLD (and the huge amount of traffic for non-existing TLD) but a big TLD has less caching before it. As of today, all the instances of one root name server, L-root, receive together around 20 kq/s. While most of it is junk (errors on the TLD name), it gives an idea of the amount of big data which pours into name servers.

A rogue DHCP server can send you to a rogue resolver. Most of the times, it seems to be done to divert traffic, by providing lies for some domain names. But it could be used just to capture the traffic and gather information about you. Same thing for malwares like DNSchanger which changes the resolver in the machine's configuration.

Many research papers about malware detection use DNS traffic to detect "abnormal" bahaviour that can be traced back to the activity of malware on infected machines. Yes, this reasearch was done for the good but, technically, it is a privacy attack and it demonstrates the power of the observation of DNS traffic. See , and .

To our knowledge, there are no specific privacy laws for DNS data. Interpreting general privacy laws like [data-protection-directive] (European Union) in the context of DNS traffic data is not an easy task and it seems there is no court precedent here.

We mention here only the solutions that could be deployed in the current Internet. Disruptive solutions, like replacing the DNS with a completely new resolution protocol, are interesting but are kept for a future work. Remember that the focus of this document is on describing the threats, not in detailing solutions. This section is therefore non-normative and is NOT a technical specification of solutions.

To defeat an eavesdropper, there is only one solution: encryption. But, from the end user point of view, even if you check that your communication between your stub resolver and the resolver is encrypted, you have no way to ensure that the communication between the resolver and the autoritative name servers will be. There are two different cases, communication between the stub resolver and the resolver (no caching but only two parties so solutions which rely on an agreement may work) and communication between the resolver and the authoritative servers (less data because of caching, but many parties involved, so any solution has to scale well). Encrypting the "last mile", between the user's stub resolver and the resolver may be sufficient since the biggest danger for privacy is between the stub resolver and the resolver, because there is no caching involved there. The only encryption mechanism available for DNS which is today an IETF standard is IPsec in ESP mode. It's deployment in the wide Internet is very limited, for reasons which are out of scope here. Still, it may be a solution for "the last mile" and, indeed, many VPN solutions use it this way, encrypting the whole traffic, including DNS to the safe resolver. In the IETF standards, a possible candidate could be DTLS . It enjoyed very little actual deployment and its interaction with the DNS has never been considered, studied or of course implemented. There are also non standard encryption techniques like DNScrypt for the stub resolver <-> resolver communication or DNScurve for the resolver <-> authoritative server communication. It seems today that the possibility of massive encryption of DNS traffic is very remote. Another solution would be to use more TCP for the queries, together with TLS . DNS can run over TCP and it provides a good way to leverage the software and experience of the TLS world. There have been discussions to use more TCP for the DNS, in light of reflection attacks (based on the spoofing of the source IP address, which is much more difficult with TCP). For instance, a stub resolver could open a TCP connection with the resolver at startup and keep it open to send queries and receive responses. The server would of course be free to tear down these connections at will (when it is under stress, for instance) and the client could reestablish them when necessary. Remember that TLS sessions can survive TCP connections so there is no need to restart the TLS negociation each time. This DNS-over-TLS-over-TCP is already implemented in the Unbound resolver. It is safe only if pipelining multiple questions over the same channel. Name compression should also be disabled, or CRIME-style attacks can apply [crime]. Encryption alone does not guarantee perfect privacy, because of the available metadata. For instance, the size of questions and responses, even encrypted, provide hints about what queries have been sent. (DNScrypt uses random-length padding, and a 64 bytes block size, to limit this risk, but this raises other issues, for instance during amplification attacks.) Observing the periodicity of encrypted questions/responses also discloses the TTL, which is yet another hint about the queries. Non-cached responses are disclosing the RTT between the resolver and authoritative servers. This is a very useful indication to guess where authoritative servers are located. Web pages are made of many resources, leading to multiple requests, whose number and timing fingerprint which web site is being browsed. So, observing encrypted traffic is not enough to recover any plaintext queries, but is enough to answer the question “is one of my employees browsing Facebook?”.Finally, attackers can perform a denial-of-service attack on possible targets, check if this makes a difference on the encrypted traffic they observe, and infer what a query was. A possible mitigation is, for clients to send “null” queries in addition to real queries. These null queries, and the related responses can have lengths and timings computed so that is it impossible for a passive attacker to distinguish these from real queries. These null queries can also help to prevent attackers from guessing what kind of connection (radio, Wifi, cable, satellite…) the user is using by measure the latency between the client and the resolver. Of course, they have a cost (more traffic) which may be unacceptable in some contexts.

It does not seem there is a possible solution against a leaky resolver. A resolver has to see the entire DNS traffic in clear. The best approach to limit the problem is to have local resolvers whose caching will limit the leak. Local networks should have a local caching resolver (even if it forwards the unanswered questions to a forwarder) and individual laptop can have its very own resolver, too.

A possible solution would be to minimize the amount of data sent from the resolver. When a resolver receives the query "What is the AAAA record for www.example.com?", it sends to the root (assuming a cold resolver, whose cache is empty) the very same question. Sending "What are the NS records for .com?" would be sufficient (since it will be the answer from the root anyway). To do so would be compatible with the current DNS system and therefore could be deployable, since it is an unilateral change to the resolvers. To do so, the resolver needs to know the zone cut . There is not a zone cut at every label boundary. If we take the name www.foo.bar.example, it is possible that there is a zone cut between "foo" and "bar" but not between "bar" and "example". So, assuming the resolver already knows the name servers of .example, when it receives the query "What is the AAAA record of www.foo.bar.example", it does not always know if the request should be sent to the name servers of bar.example or to those of example. suggest an algorithm to find the zone cut, so resolvers may try it. Note that DNSSEC-validating resolvers already have access to this information, since they have to find the zone cut (the DNSKEY record set is just below, the DS record set just above). It can be noted that minimizing the amount of data sent also partially addresses the case of a wire sniffer. Standards lovers should not that the behaviour suggested here (minimizing the amount of data sent) is NOT forbidden by the (section 5.3.3) or (section 7.2). Sending the full qname to the authoritative name server is a tradition, not a protocol MUST. Another note is that the answer to the NS query, unlike the referral sent when the question is a full qname, is in the Answer section, not in the Authoritative section. It has probably no practical consequences.

Traditional security measures (do not let malware change the system configuration) are of course a must. A protection againt rogue servers announced by DHCP could be to have a local resolver, and to always use it, ignoring DHCP.

Hey, man, the entire document is about security!

Thanks to Olaf Kolkman, Francis Dupont, Ondrej Sury and Frank Denis for the interesting discussions.