Filename: 246-merge-hsdir-and-intro.txt
Title: Merging Hidden Service Directories and Introduction Points
Author: John Brooks, George Kadianakis
Created: 2015-07-12
Status: Rejected

Change history:

   18-Jan-2016  Changed status to "Needs-Research" after discussion in email
                thread [1].

1. Overview and Motivation

   This document describes a modification to proposal 224 ("Next-Generation
   Hidden Services in Tor"), which simplifies and improves the architecture by
   combining hidden service directories and introduction points at the same

   A reader will want to be familiar with the existing hidden service design,
   and with the changes in proposal 224. If accepted, this proposal should be
   combined with proposal 224 to make a superseding specification.

1.1. Overview

   In the existing hidden service design and proposal 224, there are three
   distinct steps building a connection: fetching the descriptor from a
   directory, contacting an introduction point listed in the descriptor, and
   rendezvous as specified during the introduction. The hidden service
   directories are selected algorithmically, and introduction points are
   selected at random by the service.

   We propose to combine the responsibilities of the introduction point and
   hidden service directory. The list of introduction points responsible for a
   service will be selected using the algorithm specified for HSDirs [proposal
   224, section 2.2.3]. The service builds a long-term introduction circuit to
   each of these, identified by its blinded public key. Clients can calculate
   the same set of relays, build an introduction circuit, retrieve the
   ephemeral keys, and proceed with sending an introduction to the service in
   the same ways as before.

1.2. Benefits over proposal 224

   With this change, client connections are made more efficient by needing only
   two circuits (for introduction and rendezvous), instead of the three needed
   previously, and need to contact fewer relays. Clients also no longer cache
   descriptors, which substantially simplifies code and removes a common source
   of bugs and reliability issues.

   Hidden services are able to stay online by simply maintaining their
   introduction circuits; there is no longer a need to periodically update
   descriptors. This reduces network load and traffic fingerprinting
   opportunities for a hidden service.

   The number and churn of relays a hidden service depends on is also reduced.
   In particular, prior hidden service designs may frequently choose new
   introduction points, and each of these has an opportunity to observe the
   popularity or connection behavior of clients.

1.3. Other effects on proposal 224

   An adversarial introduction point is not significantly more capable than a
   hidden service directory under proposal 224. The differences are:

     1. The introduction point maintains a long-lived circuit with the service
     2. The introduction point can break that circuit and cause the service to
        rebuild it

   See section 4 ("Discussion") for other impacts and open discussion

2. Specification

2.1. Picking introduction points for a service

   Instead of picking HSDirs, hidden services pick their introduction points
   using the same algorithm as defined in proposal 224 section 2.2 [HASHRING].

   To be used as an introduction point, a relay must have the Stable flag in
   the consensus and an uptime of at least twice the shared random period
   defined in proposal 224 section 2.3.

   This also specifies the lifetime of introduction points, since they will be
   rotated with the change of time period and shared randomness.

2.2. Hidden service sets up introduction points

   After a hidden service has picked its intro points, it needs to establish
   long-term introduction circuits to them and also send them an encrypted
   descriptor that should be forwarded to potential clients. The descriptor
   contains a service key that should be used by clients to encrypt the
   INTRODUCE1 cell that will be sent to the hidden service. The encrypted parts
   of the descriptor are encrypted with the symmetric keys specified in prop224
   section [ENCRYPTED-DATA].

2.2.1. Hidden service uploads a descriptor

   Services post a descriptor by opening a directory stream with BEGIN_DIR, and
   sending a HTTP POST request as described in proposal 224, section 2.2.4.

   The relay must verify the signatures of the descriptor, and check whether it
   is responsible for that blinded public key in the hash ring. Relays should
   connect the descriptor to the circuit used to upload it, which will be
   repurposed as the service introduction circuit. The descriptor does not need
   to be cached by the introduction point after that introduction circuit has

   It is unexpected and invalid to send more than one descriptor on the same
   introduction circuit.

2.2.2. Descriptor format

   The format for the hidden service descriptor is as described in proposal 224
   sections 2.4 and 2.5, with the following modifications:

       * The "revision-counter" field is removed
       * The introduction-point section is removed
       * The "auth-key" field is removed
       * The "enc-key legacy" field is removed
       * The "enc-key ntor" field must be specified exactly once per descriptor

   Unlike previous versions, the descriptor does not encode the entire list of
   introduction points. The descriptor only contains a key for the particular
   introduction point it was sent to.


   When a hidden service is establishing a new introduction point, it sends the
   ESTABLISH_INTRO cell, which is formatted as described by proposal 224
   section 3.1.1, except for the following:

   The AUTH_KEY_TYPE value 02 is changed to:

     [02] -- Signing key certificate cross-certified with the blinded key, in
             the same format as in the hidden service descriptor.

   In this case, SIG is a signature of the cell with the signing key specified
   in AUTH_KEY. The relay must verify this signature, as well as the
   certification with the blinded key. The relay should also verify that it has
   received a valid descriptor with this blinded key.

   [XXX: Other options include putting only the blinded key, or only the
   signing key in this cell. In either of these cases, we must look up the
   descriptor to fully validate the contents, but we require the descriptor
   to be present anyway. -special]

   [XXX: What happens with the MAINT_INTRO process defined in proposal 224
   section 3.1.3? -special]

2.3. Client connection to a service

   A client that wants to connect to a hidden service should first calculate
   the responsible introduction points for the onion address as described in
   section 2.1 above.

   The client chooses one introduction point at random, builds a circuit, and
   fetches the descriptor. Once it has received, verified, and decrypted the
   descriptor, the client can use the same circuit to send the INTRODUCE1 cell.

2.3.1. Client requests a descriptor

   Clients can request a descriptor by opening a directory stream with
   BEGIN_DIR, and sending a HTTP GET request as described in proposal 224,
   section 2.2.4.

   The client must verify the signatures of the descriptor, and decrypt the
   encrypted portion to access the "enc-key". This key is used to encrypt the
   contents of the INTRODUCE1 cell to the service.

   Because the descriptor is specific to each introduction point, client-side
   descriptor caching changes significantly. There is little point in caching
   these descriptors, because they are inexpensive to request and will always
   be available when a service-side introduction circuit is available. A client
   that does caching must be prepared to handle INTRODUCE1 failures due to
   rotated keys.

2.3.2. Client sends INTRODUCE1

   After requesting the descriptor, the client can use the same circuit to send
   an INTRODUCE1 cell, which is forwarded to the service and begins the
   rendezvous process.

   The INTRODUCE1 cell is the same as proposal 224 section 3.2.1, except that
   the AUTH_KEYID is the blinded public key, instead of the now-removed
   introduction point authentication key.

   The relay must permit this circuit to change purpose from the directory
   request to a client or server introduction.

3. Other changes to proposal 224

3.1. Removing proposal 224 legacy relay support

   Proposal 224 defines a process for using legacy relays as introduction
   points; see section 3.1.2 [LEGACY_EST_INTRO], and 3.2.3 [LEGACY-INTRODUCE1].
   With the changes to the introduction point in this proposals, it's no longer
   possible to maintain support for legacy introduction points.

   These sections of proposal 224 are removed, along with other references to
   legacy introduction points and RSA introduction point keys. We will need to
   handle the migration process to ensure that sufficient relays are available
   as introduction points. See the discussion in section 4.1 for more details.

3.2. Removing the "introduction point authentication key"

   The "introduction point authentication key" defined in proposal 224 is
   removed.  The "descriptor signing key" is used to sign descriptors and the
   ESTABLISH_INTRO2 cell. Descriptors are unique for each introduction point,
   and there is no point in generating a new key used only to sign the

4. Discussion

4.1. No backwards compatibility with legacy relays

   By changing the introduction procedure in such a way, we are unable to
   maintain backwards compatibility. That is, hidden services will be unable to
   use old relays as their introduction points, and similarly clients will be
   unable to introduce through old relays.

   To maintain an adequate anonymity set of intro points, clients and hidden
   services should perform this introduction method only after most relays have
   upgraded. For this reason we introduce the consensus parameter
   HSMergedIntroduction which controls whether hidden services should perform
   this merged introduction or fall back to the old one.

   [XXX: Do we? This sounds like we have to implement both in the client, which
   I thought we wanted to avoid. An alternative is to make sure that the intro
   point side is done early enough, and that clients know not to rely on the
   security of 224 services until enough relays are upgraded and the
   implementation is done. -special]

4.2. Restriction on the number of intro points and impact on load balancing

   One drawback of this proposal is that the number of introduction points of a
   hidden service is now a constant global parameter. Hence, a hidden service
   can no longer adjust how many introduction points it uses, or select the
   nodes that will serve as its introduction points.

   While bad, we don't consider this a major drawback since we don't believe
   that introduction points are a significant bottleneck on hidden services

   However, our system significantly impacts the way some load balancing
   schemes for hidden services work. For example, onionbalance is a third-party
   application that manages the introduction points of a hidden service in a
   way that allows traffic load-balancing.  This is achieved by compiling a
   master descriptor that mixes and matches the introduction points of
   underlying hidden service instances.

   With our system there are no descriptors that onionbalance can use to mix
   and match introduction points. A variant of the onionbalance idea that could
   work with our system would involve onionbalance starting a hidden service,
   not establishing any intro points, and then ordering the underlying hidden
   service load-balancing instances to establish intro points to all the right
   introduction points.

4.3. Behavior when introduction points go offline or misbehave

   In this new system, it's the Tor network that decides which relays should be
   used as the intro points of a hidden service for every time period. This
   means, that a hidden service is forced to use those relays as intro points
   if it wants clients to connect to it.

   This brings up the topic of what should happen when the designated relays go
   offline or refuse connections. Our behavior here should block guard
   discovery attacks (as in #8239) while allowing maximum reachability for

   We should also make sure that an adversary cannot manipulate the hash ring
   in such a way that forces us to rotate introduction points quickly. This is
   enforced by the uptime check that is necessary for acquiring the HSDir flag

   For this reason we propose the following rules:

    - After every consensus and when the blinded public key changes as a result
      of the time period, hidden services need to recalculate their
      introduction points and adjust themselves by establishing intro points to
      the new relays.

    - When an introduction point goes offline or drops connections, we attempt
      to re-establish to it INTRO_RETRIES times per consensus. If the intro
      point failed more than INTRO_RETRIES times for a consensus period, we
      abandon it and stay with one less intro point.

      If a new consensus is released and that relay is still listed as online,
      then we reset our retry counter and start trying again.

   [XXX: Is this crazy? -asn]
   [XXX: INTRO_RETRIES = 3? -asn]

4.4. Defining constants; how many introduction points for a service?

   We keep the same intro point configuration as in proposal 224. That is, each
   hidden service uses 6 relays and keeps them for a whole time period.

   [XXX: Are these good constants? We don't have a chance to change them
   in the future!! -asn]
   [XXX: 224 makes them consensus parameters, which we can keep, but they
   can still only be changed on a network-wide basis. -special]


[1] :