Next step for DPRIVE: resolver-to-auth link

To improve the privacy of the DNS user (), the standard solution is to encrypt the requests with TLS (). Just encrypting, without authenticating the remote server, leaves the user's privacy vulnerable to active man-in-the-middle attacks. and describe how to authenticate the DNS resolver, in the stub-to-resolver link. We have currently no standard way to authenticate the authoritative name server, in the resolver-to-auth link. The two cases are quite different: a stub resolver has only a few resolvers, and there is typically a pre-existing relationship. But a resolver speaks to many authoritative name servers, without any prior relationship. This means that, for instance, having a static key for the resolver makes sense while it would be clearly unrealistic for the authoritative server. Another difference is that resolvers are typically known by IP address (obtained by DHCP or manual configuration) while authoritative name servers are known by name (obtained from zone delegation). This makes things easier for techniques similar to DANE: the manager of the ns1.example.net name server can always add a TLSA record under example.net while she may have problems modifying the zone under in-addr.arpa or ip6.arpa. (On the other hand, the resolver knows also the IP address of the authoritative server, so it can authenticate based on this, the fourth and fifth solutions in .) The original charter of the DPRIVE working group, in force at the time of this draft, says "The primary focus of this Working Group is to develop mechanisms that provide confidentiality between DNS Clients and Iterative Resolvers" and adds "but it may also later consider mechanisms that provide confidentiality between Iterative Resolvers and Authoritative Servers". This document is here for this second step, "between Iterative Resolvers and Authoritative Servers". It will probably require a rechartering of the group.

We can express the problem this way: if we want to use TLS-over-DNS to secure the link between the resolver and the authoritative server, it would be important to have a standard way to authenticate the authoritative server. Basically, the client will get the public key of the server in the TLS session, how will it know that this key is the right one? Here is a comprehensive list of the six possible solutions to this problem. First, the two where the key (or a hash of it) is available somewhere else than in the TLS session.

We could encode the key in the authoritative server name (as in DNScurve ). Here is an example of a domain using DNScurve: the names of the authoritative name servers are a Base-32 encoded form of the server's Curve25519 public key.

% dig +short NS yp.to uz5dz39x8xk8wyq3dzn7vpt670qmvzx0zd9zg4ldwldkv6kx9ft090.ns.yp.to. uz5hjgptn63q5qlch6xlrw63tf6vhvvu6mjwn0s31buw1lhmlk14kd.ns.yp.to. uz5uu2c7j228ujjccp3ustnfmr4pgcg5ylvt16kmd0qzw7bbjgd5xq.ns.yp.to. Securely transmitting the key would therefore be a by-product of delegation. Among the limits of this solution, the length of these names limit the number of possible name servers, if we want to keep the delegation short. Also, it requires a cryptographic algorithm where keys are short (which means no RSA).

We could publish keys in the DNS, secured with DNSSEC (as in DANE ). This raises an interesting bootstrap problem: we need the key to get information privately with the DNS but we need the DNS to do so. A possible solution is to have an "unsecure" mode to retrieve the initial key material. The algorithm could be: (0)The resolver remembers the keys of the authoritative name servers (in the same way it remembers the lowest RTT among a NS RRset), (1)When the resolver needs to talk to a server (say ns1.example.net) for which it does not know the key, it does a TLSA request for _853._tcp.ns1.example.net, (2)If the resolution of this request requires to talk to the very server we search the key for, the resolver connects to this server with TLS to port 853, does not authenticate, and sends the query. This step offers no authentication ("opportunistic"?). (See also , section 9.2.) The real algorithm will need to be more complicated since there are several servers per zone. A resolver may use the knowledge of TLS authentication it has to choose an authoritative name server among a NS RRset. Another solution is for the authoritative name server to use to send all the necessary DNSSEC records over TLS.

We could use the X.509 security model ). The certificates for authoritative name servers would be signed by regular CAs, with the name of the server in the Subject Alternative Name (or may be its IP address in this field, but, as far as the author knows, few CAs issue certificats for IP addresses). One of the problems is that resolvers will probably have different sets of trusted CA so an authoritative name server will not know in advance what percentage of the resolvers may authenticate it. Of course, this technique would not work if the server used raw public keys (.

The resolver could start from the IP address of the authoritative name server, and get a key from the name in the in-addr.arpa or ip6.arpa zone. For instance, if the authoritative server will be queried at 2001:db8:42:cafe::1:53, the request will be done for 3.5.0.0.1.0.0.0.0.0.0.0.0.0.0.0.e.f.a.c.2.4.0.0.8.b.d.0.1.0.0.2.ip6.arpa. There is also a boostrapping problem here too, but since there are only five RIR that manage the reverse DNS some sort of hard-coded or semi-hard-coded setup might be possible. Unfortunately the operators for the space are often different from the operators of the DNS, so this is often not a reasonable solution administratively.

Another solution starting from the IP address of the authoritative name server would be to use CGA (). The IPv6 address encodes the public key in the lower 64-bits of the address. So we could use the IPv6 address as the public key of the servers. This only works for IPv6, doesn't have (much) cryptographic agility, and raises serious layering violation issues.

Finally, we could simply not check the keys at all, accepting anything. This would break privacy promises, when there is an active attacker, able to pose as the authoritative name server. But it is still better, privacy-wise, than the current situation where DNS requests are sent in clear text.

A resolver may use a combination of these solutions. For instance, trying PKIX authentication (it does not require an extra lookup, except may be OCSP), if it fails, search a TLSA record, if there is none, depending on the resolver's policy, accept anyway. Clearly, trying all six solutions is unrealistic so some guidance on how to combine them will be necessary. All these solutions can be improved by things like automatic key pinning (). An authoritative name server cannot know if the resolver authentified it, and how. In the future, it may be interesting to have a EDNS option to signal a successful authentication, or a failure, but this is out of scope currently.

There is currently nothing to do for IANA. The future chosen solution may require some IANA action, such as a registry.

For the time being, refer to each subsection under to have an analysis of its security.

Shane Kerr for the ideas on authentication by IP address.