Network Working Group                                         D.L. Mills
Request for Comments:  1004                       University of Delaware
                                                              April 1987

A Distributed-Protocol Authentication Scheme

Status of this Memo

   The purpose of this RFC is to focus discussion on authentication
   problems in the Internet and possible methods of solution.  The
   proposed solutions this document are not intended as standards for
   the Internet at this time.  Rather, it is hoped that a general
   consensus will emerge as to the appropriate solution to
   authentication problems, leading eventually to the adoption of
   standards.  Distribution of this memo is unlimited.

1. Introduction and Overview

   This document suggests mediated access-control and authentication
   procedures suitable for those cases when an association is to be set
   up between multiple users belonging to different trust environments,
   but running distributed protocols like the existing Exterior Gateway
   Protocol (EGP) [2], proposed Dissimilar Gateway Protocol (DGP) [3]
   and similar protocols. The proposed prcedures are evolved from those
   described by Needham and Shroeder [5], but specialized to the
   distributed, multiple-user model typical of these protocols.

   The trust model and threat environment are identical to that used by
   Kent and others [1]. An association is defined as the end-to-end
   network path between two users, where the users themselves are
   secured, but the path between them is not. The network may drop,
   duplicate or deliver messages with errors. In addition, it is
   possible that a hostile user (host or gateway) might intercept,
   modify and retransmit messages. An association is similar to the
   traditional connection, but without the usual connection requirements
   for error-free delivery.  The users of the association are sometimes
   called associates.

   The proposed procedures require each association to be assigned a
   random session key, which is provided by an authentication server
   called the Cookie Jar. The procedures are designed to permit only
   those associations sanctioned by the Cookie Jar while operating over
   arbitrary network topologies, including non-secured networks and
   broadcast-media networks, and in the presence of hostile attackers.
   However, it is not the intent of these procedures to hide the data

   (except for private keys) transmitted via these networks, but only to
   authenticate messages to avoid spoofing and replay attacks.

   The procedures are intended for distributed systems where each user i
   runs a common protocol automaton using private state variables for
   each of possibly several associations simultaneously, one for each
   user j. An association is initiated by interrogating the Cookie Jar
   for a one-time key K(i,j), which is used to encrypt the checksum
   which authenticates messages exchanged between the users. The
   initiator then communicates the key to its associate as part of a
   connection establishment procedure such as described in [3].

   The information being exchanged in this protocol model is largely
   intended to converge a distributed data base to specified (as far as
   practical) contents, and does not ordinarily require a reliable
   distribution of event occurances, other than to speed the convergence
   process. Thus, the model is intrinsically resistant to message loss
   or duplication. Where important, sequence numbers are used to reduce
   the impact of message reordering. The model assumes that associations
   between peers, once having been sanctioned, are maintained
   indefinitely.  The exception when an association is broken may be due
   to a crash, loss of connectivity or administrative action such as
   reconfiguration or rekeying. Finally, the rate of information
   exchange is specifically designed to be much less than the nominal
   capabilities of the network, in order to keep overheads low.

2. Procedures

   Each user i is assigned a public address A(i) and private key K(i) by
   an out-of-band procedure beyond the scope of this discussion. The
   address can take many forms: an autonomous system identifier [2], an
   Internet address [6] or simply an arbitrary name. However, no matter
   what form it takes, every message is presumed to carry both the
   sender and receiver addresses in its header. Each address and its
   access-control list is presumed available in a public directory
   accessable to all users, but the private key is known only to the
   user and Cookie Jar and is not disclosed in messages exchanged
   between users or between users and the Cookie Jar.

   An association between i and j is identified by the bitstring
   consisting of the catenation of the addresses A(i) and A(j), together
   with a one-time key K(i,j), in the form [A(i),A(j),K(i,j)]. Note that
   the reciprocal association [A(j),A(i),K(j,i)] is distinguished only
   by which associate calls the Cookie Jar first. It is the intent in
   the protocol model that all state variables and keys relevant to a
   previous association be erased when a new association is initiated
   and no caching (as suggested in [5]) is allowed.

   The one-time key K(i,j) is generated by the Cookie Jar upon receipt
   of a request from user i to associate with user j. The Cookie Jar has
   access to a private table of entries in the form [A(i),K(i)], where i
   ranges over the set of sanctioned users. The public directory
   includes for each A(i) a list L(i) = {j1, j2, ...} of permitted
   associates for i, which can be updated only by the Cookie Jar. The
   Cookie Jar first checks that the requested user j is in L(i), then
   rolls a random number for K(i,j) and returns this to the requestor,
   which saves it and passes it along to its associate during the
   connection establishment procedure.

   In the diagrams that follow all fields not specifically mentioned are
   unencrypted. While the natural implementation would include the
   address fields of the message header in the checksum, this raises
   significant difficulties, since they may be necessary to determine
   the route through the network itself. As will be evident below, even
   if a perpetrator could successfully tamper with the address fields in
   order to cause messages to be misdelivered, the result would not be a
   useful association.

   The checksum field is computed by a algorithm using all the bits in
   the message including the address fields in the message header, then
   is ordinarily encrypted along with the sequence-number field by an
   appropriate algorithm using the specified key, so that the intended
   receiver is assured only the intended sender could have generated it.
   In the Internet system, the natural choice for checksum is the 16-
   bit, ones-complement algorithm [6], while the natural choice for
   encryption is the DES algorithm [4] (see the discussion following for
   further consideration on these points). The detailed procedures are
   as follows:

      1. The requestor i rolls a random message ID I and sends it and
      the association specifier [A(i),A(j)] as a request to the Cookie
      Jar. The message header includes the addresses [A(i),A(C)], where
      A(C) is the address of the Cookie Jar. The following schematic
      illustrates the result:

      +-----------+-----------+
      |   A(i)    |   A(C)    |       message header
      +-----------+-----------+
      |     I     | checksum  |       message ID
      +-----------+-----------+
      |   A(i)    |   A(j)    |       assoc specifier
      +-----------+-----------+

      2. The Cookie Jar checks the access list to determine if the
      association [A(i),A(j)] is valid. If so, it rolls a random number
      K(i,j) and constructs the reply below. It checksums the message,

      encrypts the j cookie field with K(j), then encrypts it and the
      other fields indicated with K(i) and finally sends the reply:

      +-----------+-----------+
      |   A(C)    |   A(i)    |       message header
      +-----------+-----------+
      |     I     | checksum  |       message ID (encrypt K(i))
      +-----------+-----------+
      |   K(i,j)  |                   i cookie (encrypt K(i))
      +-----------+
      |   K(i,j)  |                   j cookie (encrypt K(j)K(i))
      +-----------+

      3. Upon receipt of the reply the requestor i decrypts the
      indicated fields, saves the (encrypted) j cookie field and copies
      the i cookie field to the j cookie field, so that both cookie
      fields are now the original K(i,j) provided by the Cookie Jar.
      Then it verifies the checksum and matches the message ID with its
      list of outstanding requests, retaining K(i,j) for its own use. It
      then rolls a random number X for the j cookie field (to confuse
      wiretappers) and another I' for the (initial) message ID, then
      recomputes the checksum.  Finally, it inserts the saved j cookie
      field in the i cookie field, encrypts the message ID fields with
      K(i,j) and sends the following message to its associate:

      +-----------+-----------+
      |   A(i)    |   A(j)    |       message header
      +-----------+-----------+
      |     I'    | checksum  |       message ID (encrypt K(i,j))
      +-----------+-----------+
      |  K(i,j)   |                   i cookie (encrypt K(j))
      +-----------+
      |     X     |                   j cookie (noise)
      +-----------+

      4. Upon receipt of the above message the associate j decrypts the
      i cookie field, uses it to decrypt the message ID fields and
      verifies the checksum, retaining the I' and K(i,j) for later use.
      Finally, it rolls a random number J' as its own initial message
      ID, inserts it and the checksum in the confirm message, encrypts
      the message ID fields with K(i,j) and sends the message:

      +-----------+-----------+
      |   A(j)    |   A(i)    |       message header
      +-----------+-----------+
      |     J'    | checksum  |       message ID (encrypt K(i,j))
      +-----------+-----------+

      5. Subsequent messages are all coded in the same way. As new data
      are generated the message ID is incremented, a new checksum
      computed and the message ID fields encrypted with K(i,j). The
      receiver decrypts the message ID fields with K(i,j) and discards
      the message in case of incorrect checksum or sequence number.

3. Discussion

   Since the access lists are considered public read-only, there is no
   need to validate Cookie Jar requests. A perpetrator might intercept,
   modify and replay portions of Cookie Jar replies or subsequent
   messages exchanged between the associates. However, presuming the
   perpetrator does not know the keys involved, tampered messages would
   fail the checksum test and be discarded.

   The "natural" selection of Internet checksum algorithm and DES
   encryption should be reconsidered. The Internet checksum has several
   well-known vulnerabilities, including invariance to word order and
   byte swap. In addition, the checksum field itself is only sixteen
   bits wide, so a determined perpetrator might be able to discover the
   key simply by examining all possible permutations of the checksum
   field. However, the procedures proposed herein are not intended to
   compensate for weaknesses of the checksum algorithm, since this
   vulnerability exists whether authentication is used or not. Also note
   that the encrypted fields include the sequence number as well as the
   checksum. In EGP and the proposed DGP the sequence number is a
   sixteen-bit quantity and increments for each successive message,
   which makes tampering even more difficult.

   In the intended application to EGP, DGP and similar protocols
   associations may have an indefinite lifetime, although messages may
   be sent between associates on a relatively infrequent basis.
   Therefore, every association should be rekeyed occasionally, which
   can be done by either associate simply by sending a new request to
   the Cookie Jar and following the above procedure. To protect against
   stockpiling private user keys, each user should be rekeyed
   occasionally, which is equivalent to changing passwords. The
   mechanisms for doing this are beyond the scope of this proposal.

   It is a feature of this scheme that the private user keys are not
   disclosed, except to the Cookie Jar. This is why two cookies are
   used, one for i, which only it can decrypt, and the other for j,
   decrypted first by i and then by j, which only then is valid. An
   interceptor posing as i and playing back the Cookie Jar reply to j
   will be caught, since the message will fail the checksum test. A
   perpetrator with access to both the Cookie Jar reply to i and the
   subsequent message to j will see essentially a random permutation of

   all fields. In all except the first message to the Cookie Jar, the
   checksum field is encrypted, which makes it difficult to recover the
   original contents of the cookie fields before encryption by
   exploiting the properties of the checksum algorithm itself.

   The fact that the addresses in the message headers are included in
   the checksum protects against playbacks with modified addresses.
   However, it may still be possible to destabilize an association by
   playing back unmodified messages from prior associations. There are
   several possibilities:

      1. Replays of the Cookie Jar messages 1 and 2 are unlikely to
      cause damage, since the requestor matches both the addresses and
      once-only sequence number with its list of pending requests.

      2. Replays of message 3 may cause user j to falsely assume a new
      association. User j will return message 4 encrypted with the
      assumed session key, which might be an old one or even a currently
      valid one, but with invalid sequence number. Either way, user i
      will ignore message 4.

      3. Replays of message 4 or subsequent messages are unlikely to
      cause damage, since the sequence check will eliminate them.

   The second point above represents an issue of legitimate concern,
   since a determined attacker may stockpile message 3 interceptions and
   replay them at speed. While the attack is unlikely to succeed in
   establishing a working association, it might produce frequent
   timeouts and result in denial of service. In the Needham-Shroeder
   scheme this problem is avoided by requiring an additional challenge
   involving a message sent by user j and a reply sent by user i, which
   amounts to a three-way handshake.

   However, even if a three-way handshake were used, the additional
   protocol overhead induced by a determined attacker may still result
   in denial of service; moreover, the protocol model is inherently
   resistant to poor network performance, which has the same performance
   signature as the attacker. The conclusion is that the additional
   expense and overhead of a three-way handshake is unjustified.

4. Application to EGP and DGP

   This scheme can be incorporated in the Exterior Gateway Protocol
   (EGP) [2] and Dissimilar Gateway Protocol (DGP) [3] models by adding
   the fields above to the Request and Confirm messages in a
   straightforward way. An example of how this might be done is given in
   [3]. In order to retain the correctness of the state machine, it is

   convenient to treat the Cookie Jar reply as a Start event, with the
   understanding that the Cookie Jar request represents an extrinsic
   event which evokes that response.

   The neighbor-acquisition strategy intended in the Dissimilar Gateway
   Protocol DGP follows the strategy in EGP. The stability of the EGP
   state machine, used with minor modifications by DGP, was verified by
   state simulation and discussed in an appendix to [2]. Either
   associate can send a Request command at any time, which causes both
   the sender and the receiver to reinitialize all state information and
   send a Confirm response. In DGP the Request operation involves the
   Cookie Jar transaction (messages 1 and 2) and then the Request
   command itself (message 3). In DGP the keys are reinitialized as well
   and each retransmission of a Request command is separately
   authenticated.

   In DGP the Request command (message 3) and all subsequent message
   exchanges assume the keys provided by the Cookie Jar. Use of any
   other keys results in checksum discrepancies and discarded messages.
   Thus the sender knows its command has been effected, at least at the
   time the response was sent. If either associate lost its state
   variables after that time, it would ignore subsequent messages and it
   (or its associate) would eventually time out and reinitiate the whole
   procedure.

   If both associates attempt to authenticate at the same time, they may
   wind up with the authentication sequences crossing in the network.
   Note that the Request message is self-authenticating, so that if a
   Request command is received by an associate before the Confirm
   response to an earlier Request command sent by that associate, the
   keys would be reset.  Thus when the subsequent Confirm response does
   arrive, it will be disregarded and the Request command resent
   following timeout. The race that results can only be broken when, due
   to staggered timeouts, the sequences do not cross in the network.
   This is a little more complicated than EGP and does imply that more
   attention must be paid to the timeouts.

   A reliable dis-association is a slippery concept, as example TCP and
   its closing sequences. However, the protocol model here is much less
   demanding. The usual way an EGP association is dissolved is when one
   associate sends a Cease command to the other, which then sends a
   Cease-ack response; however, this is specifically assumed a non-
   reliable transaction, with timeouts specified to break retry loops.
   In any case, a new Request command will erase all history and result
   in a new association as described above.

   Other than the above, the only way to reliably dis-associate is by
   timeout. In this protocol model the associates engage in a

   reachability protocol, which requires each to send a message to the
   other from time to time. Each associate individually times out after
   a period when no messages are heard from the other.

5. Acknowledgments

   Dan Nessett and Phil Karn both provided valuable ideas and comments
   on early drafts of this report. Steve Kent and Dennis Perry both
   provided valuable advice on its review strategy.

6. References

   [1]  Kent, S.T., "Encryption-Based Protection for Interactive
        User/Computer Communication", Proc. Fifth Data Communications
        Symposium, September 1977.

   [2]  Mills, D.L., "Exterior Gateway Protocol Formal Specification",
        DARPA Network Working Group Report RFC-904, M/A-COM Linkabit,
        April 1984.

   [3]  Mills, D.L., "Dissimilar Gateway Protocol Draft Specification",
        in preparation, University of Delaware.

   [4]  National Bureau of Standards, "Data Encryption Standard",
        Federal Information Processing Standards Publication 46, January
        1977.

   [5]  Needham, R.M., and M.D. Schroeder, "Using Encryption for
        Authentication in Large Networks of Computers", Communications
        of the ACM, Vol. 21, No. 12, pp. 993-999, December 1978.

   [6]  Postel, J., "Internet Protocol", DARPA Network Working Group
        Report RFC-791, USC Information Sciences Institute, September
        1981.