Filename: 176-revising-handshake.txt
Title: Proposed version-3 link handshake for Tor
Author: Nick Mathewson
Created: 31-Jan-2011
Status: Closed
Target: 0.2.3
Supersedes: 169

1. Overview

   I propose a (mostly) backward-compatible change to the Tor
   connection establishment protocol to avoid the use of TLS
   renegotiation, to avoid certain protocol fingerprinting attacks,
   and to make it easier to write Tor clients and servers.

   Rather than doing a TLS renegotiation to exchange certificates
   and authenticate the original handshake, this proposal takes an
   approach similar to Steven Murdoch's proposal 124 and my old
   proposal 169, and uses Tor cells to finish authenticating the
   parties' identities once the initial TLS handshake is finished.

   I discuss some alternative design choices and why I didn't make
   them in section 7; please have a quick look there before
   telling me that something is pointless or makes no sense.

   Terminological note: I use "client" or "initiator" below to mean
   the Tor instance (a client or a bridge or a relay) that initiates a
   TLS connection, and "server" or "responder" to mean the Tor
   instance (a bridge or a relay) that accepts it.

2. History and Motivation

   The _goals_ of the Tor link handshake have remained basically uniform
   since our earliest versions.  They are:

      * Provide data confidentiality, data integrity
      * Provide forward secrecy
      * Allow responder authentication or bidirectional authentication.
      * Try to look like some popular too-important-to-block-at-whim
        encryption protocol, to avoid fingerprinting and censorship.
      * Try to be implementable -- on the client side at least! --
        by as many TLS implementations as possible.

   When we added the v2 handshake, we added another goal:

      * Remain compatible with older versions of the handshake
        protocol.

   In the original Tor TLS connection handshake protocol ("V1", or
   "two-cert"), parties that wanted to authenticate provided a
   two-cert chain of X.509 certificates during the handshake setup
   phase.  Every party that wanted to authenticate sent these
   certificates.  The security properties of this protocol are just
   fine; the problem was that our behavior of sending
   two-certificate chains made Tor easy to identify.

   In the current Tor TLS connection handshake protocol ("V2", or
   "renegotiating"), the parties begin with a single certificate
   sent from the server (responder) to the client (initiator), and
   then renegotiate to a two-certs-from-each-authenticating party.
   We made this change to make Tor's handshake look like a browser
   speaking SSL to a webserver.  (See proposal 130, and
   tor-spec.txt.)  So from an observer's point of view, two parties
   performing the V2 handshake begin by making a regular TLS
   handshake with a single certificate, then renegotiate
   immediately.

   To tell whether to use the V1 or V2 handshake, the servers look
   at the list of ciphers sent by the client.  (This is ugly, but
   there's not much else in the ClientHello that they can look at.)
   If the list contains any cipher not used by the V1 protocol, the
   server sends back a single cert and expects a renegotiation.  If
   the client gets back a single cert, then it withholds its own
   certificates until the TLS renegotiation phase.

   In other words, V2-supporting initiator behavior currently looks
   like this:

      - Begin TLS negotiation with V2 cipher list; wait for
        certificate(s).
      - If we get a certificate chain:
         - Then we are using the V1 handshake.  Send our own
           certificate chain as part of this initial TLS handshake
           if we want to authenticate; otherwise, send no
           certificates.  When the handshake completes, check
           certificates.  We are now mutually authenticated.

        Otherwise, if we get just a single certificate:
         - Then we are using the V2 handshake.  Do not send any
           certificates during this handshake.
         - When the handshake is done, immediately start a TLS
           renegotiation.  During the renegotiation, expect
           a certificate chain from the server; send a certificate
           chain of our own if we want to authenticate ourselves.
         - After the renegotiation, check the certificates. Then
           send (and expect) a VERSIONS cell from the other side to
           establish the link protocol version.

   And V2-supporting responder behavior now looks like this:

      - When we get a TLS ClientHello request, look at the cipher
        list.
      - If the cipher list contains only the V1 ciphersuites:
         - Then we're doing a V1 handshake.  Send a certificate
           chain.  Expect a possible client certificate chain in
           response.
        Otherwise, if we get other ciphersuites:
         - We're using the V2 handshake.  Send back a single
           certificate and let the handshake complete.
         - Do not accept any data until the client has renegotiated.
         - When the client is renegotiating, send a certificate
           chain, and expect (possibly multiple) certificates in
           reply.
         - Check the certificates when the renegotiation is done.
           Then exchange VERSIONS cells.

   Late in 2009, researchers found a flaw in most applications' use
   of TLS renegotiation: Although TLS renegotiation does not
   reauthenticate any information exchanged before the renegotiation
   takes place, many applications were treating it as though it did,
   and assuming that data sent _before_ the renegotiation was
   authenticated with the credentials negotiated _during_ the
   renegotiation.  This problem was exacerbated by the fact that
   most TLS libraries don't actually give you an obvious good way to
   tell where the renegotiation occurred relative to the datastream.
   Tor wasn't directly affected by this vulnerability, but the
   aftermath hurts us in a few ways:

      1) OpenSSL has disabled renegotiation by default, and created
         a "yes we know what we're doing" option we need to set to
         turn it back on.  (Two options, actually: one for openssl
         0.9.8l and one for 0.9.8m and later.)

      2) Some vendors have removed all renegotiation support from
         their versions of OpenSSL entirely, forcing us to tell
         users to either replace their versions of OpenSSL or to
         link Tor against a hand-built one.

      3) Because of 1 and 2, I'd expect TLS renegotiation to become
         rarer and rarer in the wild, making our own use stand out
         more.

   Furthermore, there are other issues related to TLS and
   fingerprinting that we want to fix in any revised handshake:

      1) We should make it easier to use self-signed certs, or maybe
         even existing HTTPS certificates, for the server side
         handshake, since most non-Tor SSL handshakes use either
         self-signed certificates or CA-signed certificates.

      2) We should allow other changes in our use of TLS and in our
         certificates so as to resist fingerprinting based on how
         our certificates look.  (See proposal 179.)

3. Design

3.1. The view in the large

   Taking a cue from Steven Murdoch's proposal 124 and my old
   proposal 169, I propose that we move the work currently done by
   the TLS renegotiation step (that is, authenticating the parties
   to one another) and do it with Tor cells instead of with TLS
   alone.

   This section outlines the protocol; we go into more detail below.

   To tell the client that it can use the new cell-based
   authentication system, the server sends a "V3 certificate" during
   the initial TLS handshake.  (More on what makes a certificate
   "v3" below.)  If the client recognizes the format of the
   certificate and decides to pursue the V3 handshake, then instead
   of renegotiating immediately on completion of the initial TLS
   handshake, the client instead sends a VERSIONS cell (and the
   negotiation begins).

   So the flowchart on the server side is:

      Wait for a ClientHello.
      If the client sends a ClientHello that indicates V1:
          - Send a certificate chain.
          - When the TLS handshake is done, if the client sent us a
            certificate chain, then check it.
      If the client sends a ClientHello that indicates V2 or V3:
          - Send a self-signed certificate or a CA-signed certificate
          - When the TLS handshake is done, wait for renegotiation or data.
            - If renegotiation occurs, the client is V2: send a
              certificate chain and maybe receive one.  Check the
              certificate chain as in V1.
            - If the client sends data without renegotiating, it is
              starting the V3 handshake.  Proceed with the V3
              handshake as below.

   And the client-side flowchart is:

      - Send a ClientHello with a set of ciphers that indicates V2/V3.
      - After the handshake is done:
        - If the server sent us a certificate chain, check it: we
          are using the V1 handshake.
        - If the server sent us a single "V2 certificate", we are
          using the v2 handshake: the client begins to renegotiate
          and proceeds as before.
        - Finally, if the server sent us a "v3 certificate", we are
          doing the V3 handshake below.

   And the cell-based part of the V3 handshake, in summary, is:

    C<->S: TLS handshake where S sends a "v3 certificate"

    In TLS:

       C->S: VERSIONS cell
       S->C: VERSIONS cell, CERT cell, AUTH_CHALLENGE cell, NETINFO cell

       C->S: Optionally: CERT cell, AUTHENTICATE cell
       C->S: NETINFO cell

   A "CERTS" cell contains a set of certificates; an "AUTHENTICATE"
   cell authenticates the client to the server.  More on these
   later.

3.2. Distinguishing V2 and V3 certificates

   In the protocol outline above, we require that the client can
   distinguish between v2 certificates (that is, those sent by
   current servers) and v3 certificates.  We further require that
   existing clients will accept v3 certificates as they currently
   accept v2 certificates.

   Fortunately, current certificates have a few characteristics that
   make them fairly well-mannered as it is.  We say that a certificate
   indicates a V2-only server if ALL of the following hold:
      * The certificate is not self-signed.
      * There is no DN field set in the certificate's issuer or
        subject other than "commonName".
      * The commonNames of the issuer and subject both end with
        ".net"
      * The public modulus is at most 1024 bits long.

   Otherwise, the client should assume that the server supports the
   V3 handshake.

   To the best of my knowledge, current clients will behave properly
   on receiving non-v2 certs during the initial TLS handshake so
   long as they eventually get the correct V2 cert chain during the
   renegotiation.

   The v3 requirements are easy to meet: any certificate designed to
   resist fingerprinting will likely be self-signed, or if it's
   signed by a CA, then the issuer will surely have more DN fields
   set.  Certificates that aren't trying to resist fingerprinting
   can trivially become v3 by using a CN that doesn't end with .net,
   or using a key longer than 1024 bits.

3.3. Authenticating via Tor cells: server authentication

   Once the TLS handshake is finished, if the client renegotiates,
   then the server should go on as it does currently.

   If the client implements this proposal, however, and the server
   has shown it can understand the V3+ handshake protocol, the
   client immediately sends a VERSIONS cell to the server
   and waits to receive a VERSIONS cell in return.  We negotiate
   the Tor link protocol version _before_ we proceed with the
   negotiation, in case we need to change the authentication
   protocol in the future.

   Once either party has seen the VERSIONS cell from the other, it
   knows which version they will pick (that is, the highest version
   shared by both parties' VERSIONS cells).  All Tor instances using
   the handshake protocol described in 3.2 MUST support at least
   link protocol version 3 as described here.  If a version lower
   than 3 is negotiated with the V3 handshake in place, a Tor
   instance MUST close the connection.

   On learning the link protocol, the server then sends the client a
   CERT cell and a NETINFO cell.  If the client wants to
   authenticate to the server, it sends a CERT cell, an AUTHENTICATE
   cell, and a NETINFO cell; or it may simply send a NETINFO cell if
   it does not want to authenticate.

   The CERT cell describes the keys that a Tor instance is claiming
   to have.  It is a variable-length cell.  Its payload format is:

        N: Number of certs in cell            [1 octet]
        N times:
           CertType                           [1 octet]
           CLEN                               [2 octets]
           Certificate                        [CLEN octets]

   Any extra octets at the end of a CERT cell MUST be ignored.

     CertType values are:
        1: Link key certificate from RSA1024 identity
        2: RSA1024 Identity certificate
        3: RSA1024 AUTHENTICATE cell link certificate

   The certificate format is X509.

   To authenticate the server, the client MUST check the following:
     * The CERTS cell contains exactly one CertType 1 "Link" certificate.
     * The CERTS cell contains exactly one CertType 2 "ID" certificate.
     * Both certificates have validAfter and validUntil dates that
       are not expired.
     * The certified key in the Link certificate matches the
       link key that was used to negotiate the TLS connection.
     * The certified key in the ID certificate is a 1024-bit RSA key.
     * The certified key in the ID certificate was used to sign both
       certificates.
     * The link certificate is correctly signed with the key in the
       ID certificate
     * The ID certificate is correctly self-signed.

   If all of these conditions hold, then the client knows that it is
   connected to the server whose identity key is certified in the ID
   certificate.  If any condition does not hold, the client closes
   the connection.  If the client wanted to connect to a server with
   a different identity key, the client closes the connection.

   An AUTH_CHALLENGE cell is a variable-length cell with the following
   fields:
       Challenge [32 octets]
       N_Methods [2 octets]
       Methods   [2 * N_Methods octets]

   It is sent from the server to the client.  Clients MUST ignore
   unexpected bytes at the end of the cell.  Servers MUST generate
   every challenge using a strong RNG or PRNG.

   The Challenge field is a randomly generated string that the
   client must sign (a hash of) as part of authenticating.  The
   methods are the authentication methods that the server will
   accept.  Only one authentication method is defined right now; see
   3.4 below.

3.4. Authenticating via Tor cells: Client authentication

   A client does not need to authenticate to the server.  If it
   does not wish to, it responds to the server's valid CERT cell by
   sending a NETINFO cell: once it has gotten a valid NETINFO cell,
   the client should consider the connection open, and the
   server should consider the connection as opened by an
   unauthenticated client.

   If a client wants to authenticate, it responds to the
   AUTH_CHALLENGE cell with a CERT cell and an AUTHENTICATE cell.
   The CERT cell is as a server would send, except that instead of
   sending a CertType 1 cert for an arbitrary link certificate, the
   client sends a CertType 3 cert for an RSA AUTHENTICATE key.
   (This difference is because we allow any link key type on a TLS
   link, but the protocol described here will only work for 1024-bit
   RSA keys.  A later protocol version should extend the protocol
   here to work with non-1024-bit, non-RSA keys.)

        AuthType                              [2 octets]
        AuthLen                               [2 octets]
        Authentication                        [AuthLen octets]

   Servers MUST ignore extra bytes at the end of an AUTHENTICATE
   cell.  If AuthType is 1 (meaning "RSA-SHA256-TLSSecret"), then the
   Authentication contains the following:

       TYPE: The characters "AUTH0001" [8 octets]
       CID: A SHA256 hash of the client's RSA1024 identity key [32 octets]
       SID: A SHA256 hash of the server's RSA1024 identity key [32 octets]
       SLOG: A SHA256 hash of all bytes sent from the server to the client
         as part of the negotiation up to and including the
         AUTH_CHALLENGE cell; that is, the VERSIONS cell,
         the CERT cell, the AUTH_CHALLENGE cell, and any padding cells.
         [32 octets]
       CLOG: A SHA256 hash of all bytes sent from the client to the
         server as part of the negotiation so far; that is, the
         VERSIONS cell and the CERT cell and any padding cells. [32 octets]
       SCERT: A SHA256 hash of the server's TLS link
         certificate. [32 octets]
       TLSSECRETS: A SHA256 HMAC, using the TLS master secret as the
         secret key, of the following:
           - client_random, as sent in the TLS Client Hello
           - server_random, as sent in the TLS Server Hello
           - the NUL terminated ASCII string:
             "Tor V3 handshake TLS cross-certification"
          [32 octets]
       TIME: The time of day in seconds since the POSIX epoch. [8 octets]
       RAND: A 16 byte value, randomly chosen by the client [16 octets]
       SIG: A signature of a SHA256 hash of all the previous fields
         using the client's "Authenticate" key as presented.  (As
         always in Tor, we use OAEP-MGF1 padding; see tor-spec.txt
         section 0.3.)
          [variable length]

   To check the AUTHENTICATE cell, a server checks that all fields
   containing from TYPE through TLSSECRETS contain their unique
   correct values as described above, and then verifies the signature.
   signature.  The server MUST ignore any extra bytes in the signed
   data after the SHA256 hash.

3.5. Responding to extra cells, and other security checks.

   If the handshake is a V3 TLS handshake, both parties MUST reject
   any negotiated link version less than 3.  Both parties MUST check
   this and close the connection if it is violated.

   If the handshake is not a V3 TLS handshake, both parties MUST
   still advertise all link protocols they support in their versions
   cell.  Both parties MUST close the link if it turns out they both
   would have supported version 3 or higher, but they somehow wound
   up using a v2 or v1 handshake.  (More on this in section 6.4.)

   Either party may send a VPADDING cell at any time during the
   handshake, except as the first cell. (See proposal 184.)

   A server SHOULD NOT send any sequence of cells when starting a v3
   negotiation other than "VERSIONS, CERT, AUTH_CHALLENGE,
   NETINFO".  A client SHOULD drop a CERT, AUTH_CHALLENGE, or
   NETINFO cell that appears at any other time or out of sequence.

   A client should not begin a v3 negotiation with any sequence
   other than "VERSIONS, NETINFO" or "VERSIONS, CERT, AUTHENTICATE,
   NETINFO".   A server SHOULD drop a CERT, AUTH_CHALLENGE, or
   NETINFO cell that appears at any other time or out of sequence.

4. Numbers to assign

   We need a version number for this link protocol.  I've been
   calling it "3".

   We need to reserve command numbers for CERT, AUTH_CHALLENGE, and
   AUTHENTICATE.  I suggest that in link protocol 3 and higher, we
   reserve a separate range of commands for variable-length cells.
   See proposal 184 for more there.

5. Efficiency

   This protocol adds a round-trip step when the client sends a
   VERSIONS cell to the server, and waits for the {VERSIONS, CERT,
   NETINFO} response in turn.  (The server then waits for the
   client's {NETINFO} or {CERT, AUTHENTICATE, NETINFO} reply,
   but it would have already been waiting for the client's NETINFO,
   so that's not an additional wait.)

   This is actually fewer round-trip steps than required before for
   TLS renegotiation, so that's a win over v2.

6. Security argument

   These aren't crypto proofs, since I don't write those.  They are
   meant to be reasonably convincing.

6.1. The server is authenticated

   TLS guarantees that if the TLS handshake completes successfully,
   the client knows that it is speaking to somebody who knows the
   private key corresponding to the public link key that was used in
   the TLS handshake.

   Because this public link key is signed by the server's identity
   key in the CERT cell, the client knows that somebody who holds
   the server's private identity key says that the server's public
   link key corresponds to the server's public identity key.

   Therefore, if the crypto works, and if TLS works, and if the keys
   aren't compromised, then the client is talking to somebody who
   holds the server's private identity key.

6.2. The client is authenticated

   Once the server has checked the client's certificates, the server
   knows that somebody who knows the client's private identity key
   says that he is the one holding the private key corresponding to
   the client's presented link-authentication public key.

   Once the server has checked the signature in the AUTHENTICATE
   cell, the server knows that somebody holding the client's
   link-authentication private key signed the data in question.  By
   the standard certification argument above, the server knows that
   somebody holding the client's private identity key signed the
   data in question.

   So the server's remaining question is: am I really talking to
   somebody holding the client's identity key, or am I getting a
   replayed or MITM'd AUTHENTICATE cell that was previously sent by
   the client?

   Because the client includes a TLSSECRET component, and the
   server is able to verify it, then the answer is easy: the server
   knows for certain that it is talking to the party with whom it
   did the TLS handshake, since if somebody else generated a correct
   TLSSECRET, they would have to know the master secret of the TLS
   connection, which would require them to have broken TLS.

   Even if the protocol didn't contain the TLSSECRET component,
   the server could the client's authentication, but it's a little
   trickier.  The server knows that it is not getting a replayed
   AUTHENTICATE cell, since the cell authenticates (among other
   stuff) the server's AUTH_CHALLENGE cell, which it has never used
   before.  The server knows that it is not getting a MITM'd
   AUTHENTICATE cell, since the cell includes a hash of the server's
   link certificate, which nobody else should have been able to use
   in a successful TLS negotiation.

6.3. MITM attacks won't work any better than they do against TLS

   TLS guarantees that a man-in-the-middle attacker can't read the
   content of a successfully negotiated encrypted connection, nor
   alter the content in any way other than truncating it, unless he
   compromises the session keys or one of the key-exchange secret
   keys used to establish that connection.  Let's make sure we do at
   least that well.

   Suppose that a client Alice connects to an MITM attacker Mallory,
   thinking that she is connecting to some server Bob.  Let's assume
   that the TLS handshake between Alice and Mallory finishes
   successfully and the v3 protocol is chosen.  [If the v1 or v2
   protocol is chosen, those already resist MITM.  If the TLS
   handshake doesn't complete, then Alice isn't connected to anybody.]

   During the v3 handshake, Mallory can't convince Alice that she is
   talking to Bob, since she should not be able to produce a CERT
   cell containing a certificate chain signed by Bob's identity key
   and used to authenticate the link key that Mallory used during
   TLS.  (If Mallory used her own link key for the TLS handshake, it
   won't match anything Bob signed unless Bob is compromised.
   Mallory can't use any key that Bob _did_ produce a certificate
   for, since she doesn't know the private key.)

   Even if Alice fails to check the certificates from Bob, Mallory
   still can't convince Bob that she is really Alice.  Assuming that
   Alice's keys aren't compromised, Mallory can't send a CERT cell
   with a cert chain from Alice's identity key to a key that Mallory
   controls, so if Mallory wants to impersonate Alice's identity
   key, she can only do so by sending an AUTHENTICATE cell really
   generated by Alice.  Because Bob will check that the random bytes
   in the AUTH_CHALLENGE cell will influence the SLOG hash, Mallory
   needs to send Bob's challenge to Alice, and can't use any other
   AUTHENTICATE cell that Alice generated before.  But because the
   AUTHENTICATE cell Alice will generate will include in the SCERT
   field a hash of the link certificate used by Mallory, Bob will
   reject it as not being valid to connect to him.

6.4. Protocol downgrade attacks won't work.

   Assuming that Alice checks the certificates from Bob, she knows
   that Bob really sent her the VERSION cell that she received.

   Because the AUTHENTICATE cell from Alice includes signed hashes
   of the VERSIONS cells from Alice and Bob, Bob knows that Alice
   got the VERSIONS cell he sent and sent the VERSIONS cell that he
   received.

   But what about attempts to downgrade the protocol earlier in the
   handshake?  Here TLS comes to the rescue: because the TLS
   Finished handshake message includes an authenticated digest of
   everything previously said during the handshake, an attacker
   can't replace the client's ciphersuite list (to trigger a
   downgrade to the v1 protocol) or the server's certificate [chain]
   (to trigger a downgrade to the v1 or v2 protocol).

7. Design considerations

   I previously considered adding our own certificate format in
   order to avoid the pain associated with X509, but decided instead
   to simply use X509 since a correct Tor implementation will
   already need to have X509 code to handle the other handshake
   versions and to use TLS.

   The trickiest part of the design here is deciding what to stick
   in the AUTHENTICATE cell.  Some of it is strictly necessary, and
   some of it is left there for security margin in case my other
   security arguments fail.  Because of the CID and SID elements
   you can't use an AUTHENTICATE cell for anything other than
   authenticating a client ID to a server with an appropriate
   server ID.  The SLOG and CLOG elements are there mostly to
   authenticate the VERSIONS cells and resist downgrade attacks
   once there are two versions of this.  The presence of the
   AUTH_CHALLENGE field in the stuff authenticated in SLOG
   prevents replays and ensures that the AUTHENTICATE cell was
   really generated by somebody who is reading what the server is
   sending over the TLS connection.  The SCERT element is meant to
   prevent MITM attacks.  When the TLSSECRET field is
   used, it should prevent the use of the AUTHENTICATE cell for
   anything other than the TLS connection the client had in mind.

   A signature of the TLSSECRET element on its own should also be
   sufficient to prevent the attacks we care about.  The redundancy
   here should come in handy if I've made a mistake somewhere else in
   my analysis.

   If the client checks the server's certificates and matches them
   to the TLS connection link key before proceding with the
   handshake, then signing the contents of the AUTH_CHALLENGE cell
   would be sufficient to authenticate the client.  But implementers
   of allegedly compatible Tor clients have in the past skipped
   certificate verification steps, and I didn't want a client's
   failure to verify certificates to mean that a server couldn't
   trust that he was really talking to the client.  To prevent this,
   I added the TLS link certificate to the authenticated data: even
   if the Tor client code doesn't check any certificates, the TLS
   library code will still check that the certificate used in the
   handshake contains a link key that matches the one used in the
   handshake.

8. Open questions:

  - May we cache which certificates we've already verified?  It
    might leak in timing whether we've connected with a given server
    before, and how recently.

  - With which TLS libraries is it feasible to yoink client_random,
    server_random, and the master secret?  If the answer is "All
    free C TLS libraries", great.  If the answer is "OpenSSL only",
    not so great.

  - Should we do anything to check the timestamp in the AUTHENTICATE
    cell?

  - Can we give some way for clients to signal "I want to use the
    V3 protocol if possible, but I can't renegotiate, so don't give
    me the V2"?  Clients currently have a fair idea of server
    versions, so they could potentially do the V3 handshake with
    servers that support it, and fall back to V1 otherwise.

  - What should servers that don't have TLS renegotiation do?  For
    now, I think they should just stick with V1.  Eventually we can
    deprecate the V2 handshake as we did with the V1 handshake.
    When that happens, servers can be V3-only.