INTERNET-DRAFT G Brown draft-petke-mech-00.txt CompuServe Expires: 15-May-97 15 November 1996 Remote Passphrase Authentication Part Two: The Mechanism Status of this Memo This document is an Internet-Draft. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that other groups may also distribute working documents as Internet-Drafts. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet- Drafts as reference material or to cite them other than as "work in progress." To learn the current status of any Internet-Draft, please check the "1id-abstracts.txt" listing contained in the Internet- Drafts Shadow Directories on ftp.is.co.za (Africa), nic.nordu.net (Europe), munnari.oz.au (Pacific Rim), ds.internic.net (US East Coast), or ftp.isi.edu (US West Coast). Abstract Remote Passphrase Authentication provides a way to authenticate a user to a service by using a pass phrase over an insecure network, without revealing the pass phrase to eavesdroppers. In addition, the service need not know and does not learn the user's pass phrase, making this scheme useful in distributed environments where it would be difficult or inappropriate to trust a service with a pass phrase database or to allow the server to learn enough to masquerade as the user in a future authentication attempt. This draft is part two of a four part series and explains the mechanism behind RPA. Part one of this series (draft-petke-ext-intro-00.txt) provides an extended introduction to the problems of authentication over insecure networks. Part three (draft-petke-http-auth-scheme-00.txt) explains how to incorporate the mechanism into HTTP. Part four (draft-petke-serv-deity-protocol-00.txt) explains the protocol between the service and deity. This scheme was inspired by Dave Raggett's Mediated Digest Authentication paper. G Brown [Page 1] INTERNET-DRAFT RPA - Part Two 15 November 1996 Table of Contents 1. INTRODUCTION 2. TERMINOLOGY 3. DESIGN CRITERIA 4. THE MECHANISM 4.1 AUTHENTICATION 4.1.1 Values and their representation 4.1.2 The authentication process 4.2 REAUTHENTICATION 4.3 REAUTHENTICATION CHEATING 5. SECURITY CONSIDERATIONS 6. AUTHOR'S ADDRESS 1. Introduction In this mechanism, we'll authenticate a user to a service and vice versa. We'll use pass phrases--actually, they're 128-bit shared secrets, but we'll define a way to use textual phrases--so the goal is to prove to the service that you know your pass phrase, and vice versa. Of course, it's important not to reveal the pass phrase to an eavesdropper. It is equally important not to reveal the pass phrase to a spoofer. Furthermore, the mechanism should work even if the service does not know the user's pass phrase. In a distributed environment, with many services that wish to authenticate the same set of users, it may be difficult to make users' pass phrases available to all services. And we might prefer not to do that, if we don't completely trust the services. So, not only should the service not have to know the user's pass phrase, but the service should not learn the user's pass phrase during the authentication process. On the other hand, the mechanism should be simple enough to apply even in the traditional case where the service knows the user's pass phrase; there's no need to use a different mechanism in that case. Part one of this specification (draft-petke-ext-intro-00.txt) contains an extended introduction that explains the problem and various potential solutions and their problems, leading to this mechanism. If you find yourself asking, "Why not just...," it might G Brown [Page 2] INTERNET-DRAFT RPA - Part Two 15 November 1996 be worth reading part one to see if that explains it. However, it contains only background material, so you needn't read part one before reading the rest of this specification. 2. Terminology Throughout this specification we'll speak of a "user" communicating with a "service" that wishes to learn and authenticate the user's identity. Often, the user is a "client" and the service is a "server," but those terms refer to an implementation. The "deity" knows the users' and services' pass phrases, and the service talks to the deity during the authentication process. Although the term "authentication server" is more conventional, we call it a deity because it's got fewer syllables and the term "server" is overloaded. If the service knows the pass phrases, then it acts as its own deity, simplifying the implementation but otherwise having no effect on the mechanism. Identities exist in some "realm," and we use that term in its usual sense. We often think of a realm as being a relatively large collection of users, like compuserve.com or aol.com, but it might well consist of a small set of users, e.g., user names and pass phrases associated with an individual Web server. We allow the service to specify a set of realms, to recognize an identity in any of the realms in which it participates. 3. Design criteria This authentication mechanism is intended to meet the following criteria. * The service learns and authenticates the user's identity. * The user learns and authenticates the service's identity. * The mechanism does not use public-key technology. * The mechanism does not use encryption. (By encryption, we're referring to reversible encryption, the ability to encrypt something and later decrypt it. By avoiding encryption, we avoid restrictions on exportability.) * The mechanism is based on shared secrets: "pass phrases," although they can be arbitrary bit patterns rather than text. * Neither the user nor the service needs to know the other's pass phrase. G Brown [Page 3] INTERNET-DRAFT RPA - Part Two 15 November 1996 * Neither the user nor the service nor eavesdroppers will learn the other's pass phrase. However, if the pass phrase is based on text, it's important to choose a "good" pass phrase to avoid a dictionary attack. * The mechanism is reasonably easy to implement in clients and does not require the client to communicate with a third party nor to a possess a reliable clock. * The mechanism derives a shared secret that may be used as a session key for subsequent authentication. * The mechanism may be incorporated into almost any protocol. In other words, the mechanism is not designed around a protocol; the protocol is designed around the mechanism. But the mechanism must be suitable for incorporation into protocols like HTTP. * The mechanism provides the ability to accept an identity in any of a set of realms in which the user and service are members. 4. The Mechanism This authentication mechanism consists of three related processes: authentication, reauthentication, and reauthentication cheating. Authentication is the fundamental process by which a user and a service mutually authenticate one another within one of a set of realms, without revealing their pass phrases to one another. Reauthentication is a process by which a user and service, having recently authenticated one another, may again authenticate one another. They could, of course, simply repeat the authentication process, but that requires interaction with an authentication deity. The reauthentication process is faster, requiring no communication with a third party. Reauthentication is useful when multiple connections between the user and service are established, whether sequential as in HTTP or simultaneous. Each connection must be authenticated, but the reauthentication process provides a shortcut. Reauthentication cheating is a further optimization for HTTP, a protocol that is quite unfriendly to challenge-response mechanisms. Reauthentication cheating can be performed in parallel with an HTTP transaction. True reauthentication is just as simple, but requires two sequential requests because of the characteristics of HTTP. By using reauthentication cheating, we create a "one-way" handshake. G Brown [Page 4] INTERNET-DRAFT RPA - Part Two 15 November 1996 4.1 Authentication There are three parties involved in the authentication process: * the user; * the service; and * the authentication deity. Each user has a name and a pass phrase in some realm of interest. Similarly, each service has a name and a pass phrase in that realm. The pass phrase isn't really text; it's a 128-bit (16-octet) string of bits. However, it's often useful to use pass phrases in the conventional, textual sense, so we define a procedure for converting a textual phrase to the 128-bit value used by the authentication mechanism. If such a pass phrase is poorly chosen, it will be subject to dictionary attack, and that's why we never use the word password in this specification (well, except in this sentence)--use a phrase, not a word. The service may specify a list of realms, and the user chooses one in which he has an identity. Thus, a service is not restricted to authenticating identities in a single realm. The service must possess a name and pass phrase in all realms it lists. Each realm has an authentication deity, which knows the names and pass phrases of its members. It's the service's responsibility to know how to locate an authentication deity for each realm; the user never communicates directly with an authentication deity. If the service knows the user's pass phrase, it performs the role of the authentication deity itself, but this does not affect the mechanism. 4.1.1 Values and their representation Following is a glossary of the values involved in the authentication process; we'll use these symbols in the following explanation. As--Authentication deity's response to service; proves user's identity Au--Authentication deity's response to user; proves service's identity Cs--Challenge from service Cu--Challenge from user G Brown [Page 5] INTERNET-DRAFT RPA - Part Two 15 November 1996 Kus--Session key for user and service Kuss--Session key obscured so visible only to service Kusu--Session key obscured so visible only to user Nr--Realm name Ns--Service name Nu--User name Ps--Service's pass phrase, a 128-bit value Pu--User's pass phrase, a 128-bit value Rs--Service's response to challenge (during authentication process, goes to authentication deity; during reauthentication, goes to user) Ru--User's response to challenge (during authentication process, goes via service to authentication deity; during reauthentication, goes to service) Ts--Service's time stamp Z--Padding consisting of 48 octets (384 bits) with all bits set to zero +--Concatenation of octet strings xor--Bitwise exclusive or Bit patterns for each value must be specified. Imagine, for example, that one implementation uses ASCII, another EBCDIC, and another Unicode for the user name. Or one implementation converts the name to lowercase, another to all caps. Each would generate a different result for the same calculation, and authentication would fail. Should we leave such details to the underlying protocol? We could, but that would make the service-to-deity protocol dependent on the user-to-service protocol, so we couldn't have a single deity for each realm. If we specify the bit patterns, we can allow any mixture of user-to-service and service-to-deity protocols to operate on the same data. Therefore, we adopt the following conventions. Text strings are represented in the Unicode character set, in big-endian byte order, without a trailing null character. Note that ASCII can be converted to ISO 8859-1 by prefixing a single 0 bit, and ISO 8859-1 can be converted to Unicode by prefixing eight 0 bits. G Brown [Page 6] INTERNET-DRAFT RPA - Part Two 15 November 1996 Each 16-bit Unicode character is stored in two octets, with its high-order 8 bits in the first octet. Representation of characters with multiple encodings is for further study. For example, e-acute has more than one representation. The form that uses combining characters, in character-code order, is probably the most logical. Note, by the way, that this specification refers only to values used in the authentication calculations, not the underlying protocol. For example, it's quite reasonable for a protocol to use ASCII for user names, if that character set is adequate. Those ASCII characters must be converted to Unicode before using them in authentication calculations, but the protocol need not transmit Unicode characters. * Names--Nr, Ns, Nu--are converted to lowercase Unicode. Note that there is no trailing null character. * Challenges--Cs, Cu--are arbitrary strings of octets, not text. They may contain any bit patterns, including nulls, and must be at least eight octets in length. * The time stamp--Ts--is the ISO 646 (ASCII) textual representation of the current universal time--UTC--in exactly 14 octets, using 24-hour time, with leading zeroes: 19950805011344. * Pass phrases--Ps, Pu--are 16-octet quantities that contain arbitrary bit patterns, including nulls. If the pass phrase is based on a textual phrase, the textual phrase is converted to a 16-octet quantity by the following process. * Convert the text string to a sequence of characters in either the Unicode or ISO 8859-1 character sets, as appropriate for the realm. * Convert each character to its lowercase equivalent, or its uppercase equivalent, or leave it alone, as appropriate for the realm. * Store the sequence of characters in an octet stream, with each Unicode character in two consecutive octets in big-endian order, or each ISO 8859-1 character in one octet. Do not append a trailing null character. * Take the MD5 digest of the resulting string of octets. The result is the 128-bit value to use in the authentication calculations. A realm will specify which of the preceding options--character set, case conversion, and hash function--it uses for the text-to-128-bit value transformation; the defaults are Unicode, convert to lowercase, G Brown [Page 7] INTERNET-DRAFT RPA - Part Two 15 November 1996 and MD5. More options might be added in the future. The user-service protocol should be designed to convey the appropriate options for each realm from the service to the user, if other than the defaults are to be supported, to avoid requiring the (human) user to manually configure software. 4.1.2 The authentication process Here we describe the individual steps. Taken literally, one might envision many messages between the service and deity, but an actual implementation within a protocol combines steps. [Perhaps we should add a "sample protocol" section showing a three-way handshake version.] For example, "The user sends a random challenge" is shown as a separate step for clarity, but it needn't be a separate message to the service, nor must it be sent at the point shown--if it makes sense in the underlying protocol, the user's challenge might be included with the user's response to the service. * The service supplies a sequence of realms, with the service's name in each realm, to the user. For example, foo@compuserve.com bar@aol.com means "Please identify yourself with a CIS user ID. If you don't have one, your AOL ID will do." The service indicates its realm preferences in most-preferred to least-preferred order; by specifying only one realm, the service requires identification in that realm. * The user chooses a realm, Nr, and gives it and his name in that realm, Nu, to the service. That, in turn, determines Ns, the service's name in that realm. Note that a protocol might allow the service to include a null realm name, meaning "I'll accept you as an anonymous user if you wish." The user might make this choice by supplying a null name; the process stops here, and no authentication is performed. * The service transmits a random challenge, Cs, and a time stamp, Ts. The challenges are random values that make each authentication unique. The time stamp is, in effect, a third challenge, which the deity will ensure is recent. The user may examine it, but most users lack an accurate source of universal time, so most users will treat it as an opaque value. * The user sends a random challenge, Cu. G Brown [Page 8] INTERNET-DRAFT RPA - Part Two 15 November 1996 * The user calculates a response, Ru: Ru = MD5(Pu + Z + Nu + Ns + Nr + Cu + Cs + Ts + Pu) and sends it to the service. Only the real user can generate the correct response, because it depends on the user's pass phrase, Pu. No one can determine the user's pass phrase from a captured response, because it's generated by a one-way function, although there is the risk of a dictionary attack if Pu is based on a poorly chosen pass phrase. * The service calculates a response, Rs: Rs = MD5(Ps + Z + Nu + Ns + Nr + Cu + Cs + Ts + Ru + Ps) This response is not sent to the user; it would do no harm if the user saw it, but the user won't need it. * The service sends a request to the authentication deity for the realm in question. The request contains - The realm name, Nr (included so the same deity can serve more than one realm) - The user's name, Nu - The service's name, Ns - The user's challenge, Cu - The service's challenge, Cs - The time stamp, Ts - The user's response, Ru - The service's response, Rs * The deity verifies the time stamp per previously agreed upon criteria. In some applications, one might require it within a few minutes; in others, one might want to allow 25 hours to eliminate problems of misconfigured time zones. Beware of overzealousness, though; this time stamp went from the service to the user, then back to the service, then to the deity, perhaps with human interaction--typing a pass phrase--introducing further delay. The deity might implement a replay cache. * The deity uses Nr, Ns, and Nu to look up the user's and service's pass phrases. G Brown [Page 9] INTERNET-DRAFT RPA - Part Two 15 November 1996 * The deity uses the values in the request, plus the service's pass phrase, Ps, to verify Rs. If it is incorrect, the deity returns a negative response; this request apparently did not come from a legitimate service. * Having verified the requesting service's identity, the deity uses the values in the request, plus the user's pass phrase, Pu, to verify Ru. If it is incorrect, the deity returns a failure response to the service; the user does not know the correct pass phrase. [Think about the effects of replaying the authentication request. I think the answer is that there's no problem because it reveals no new information.] * Having verified both the user's and service's identity, the deity creates a random, 128-bit session key, Kus, for use by the user and service. They might use it for session encryption; in addition, it will be used in the reauthentication process described later. * The deity generates two obscured copies of the session key: - Kuss = Kus xor MD5(Ps + Z + Ns + Nu + Nr + Cs + Cu + Ts + Ps) - Kusu = Kus xor MD5(Pu + Z + Ns + Nu + Nr + Cs + Cu + Ts + Pu) The obscuring masks resemble Ru and Rs, but differ, of course, so an eavesdropper cannot recover Kus. * The deity generates a pair of authentication "proofs": - Au = MD5(Pu + Z + Ns + Nu + Nr + Kusu + Cs + Cu + Ts + Kus + Pu) - As = MD5(Ps + Z + Ns + Nu + Nr + Kuss + Cs + Cu + Ts + Kus + M + Ps) Here "M" is the message transmitted from the deity to the service; it is included in the calculation to authenticate the response to the service. Refer to part four of this specification (draft-petke-serv-deity-protocol-00.txt) for more details. * The deity sends the four values Kuss, Kusu, As, and Au to the service. * The service extracts its copy of the session key from Kuss by calculating the obscuring mask value and XORing. (The service can determine the user's key-obscuring value by calculating Kus xor Kusu; and if the user sees Kuss, it can do likewise. But the obscuring masks reveal nothing.) G Brown [Page 10] INTERNET-DRAFT RPA - Part Two 15 November 1996 * The service verifies As by performing the same calculation and comparing the result. If it matches, the service knows that someone who knows its pass phrase--the deity--replied, having verified that the user is who he claims to be. * The service forwards Kusu and Au to the user. * The user extracts its copy of the session key from Kusu by calculating the mask value and XORing. * The user verifies Au by computing it and comparing. If it matches, the user knows that someone who knows his pass phrase--the deity--replied, having verified that the service is who it claims to be. Of course, if the service itself knows the user's pass phrase, it can assert any service identity; but this is the case where the service is trusted and acts as its own deity. Now the user and service are confident of each others' identities, and the two parties share a session key that they may use for encryption, if they so choose. [Perhaps we should add another value to the authentication calculations, opaque to the mechanism, provided by the protocol in which this mechanism is embedded. This value would, of course, have to be added to the service-to-deity protocol, and its generation and interpretation would be up to the lower-level protocol. For example, HTTP might choose to include the Web server's IP address and, perhaps, the URL in the authentication calculations, making it harder to do a man-in-the-middle attack. (Of course, that problem cannot be completely solved without using the session key to authenticate data, which is a protocol issue outside the scope of this mechanism.)] 4.2 Reauthentication Reauthentication is a process by which a user and service, having recently authenticated each other, may again mutually authenticate without talking to a deity. This is useful with protocols like HTTP, which involve a sequence of connections that must be independently authenticated. It's also useful with parallel connections--imagine a scheme in which a user and service are connected, and wish to establish a second connection. To reauthenticate one another, the user and service prove to each other that they both possess a secret 128-bit key--the session key, Kus, derived during the authentication process. The reauthentication process is essentially an ordinary challenge-response mechanism in which the session key is used as a pass phrase. G Brown [Page 11] INTERNET-DRAFT RPA - Part Two 15 November 1996 * The service sends a challenge, Cs, to the user. * The user sends a challenge, Cu, to the service. * The user calculates Ru = MD5(Kus + Z + Ns + Nu + Nr + Cs + Cu + Kus) and sends it to the service. * The service verifies the result. If correct, it calculates Rs = MD5(Kus + Z + Nu + Ns + Nr + Cu + Cs + Kus) and sends it to the user. Both responses involve the same set of values, but they're used in a different order, so the responses are different. * The user verifies the result. 4.3 Reauthentication cheating In HTTP, one can shortcut the reauthentication process by cheating, for an increase in efficiency. A naive approach allows the user to repeat its authentication data, presumably in the form of an Authorization header. If the service recognizes the same Authorization header, it presumes that it's talking to the previously authenticated user; essentially, we pretend that we reauthenticated with the same challenges. But this approach is vulnerable to replay attacks during the period of time the service considers the data valid. The service can check the user's IP address to reduce the risk, but IP addresses mean surprisingly little. Even neglecting address spoofing, multiple users share an IP address when they're on the same host or routed through a proxy or SOCKS server. There's a better solution. We begin by noting why it's desirable--from an efficiency, not security, point of view--to allow the Authorization header to be replayed. To embed a challenge-response mechanism in HTTP, we require at least two HTTP transactions for authentication, because we cannot send a challenge and receive a response in one HTTP transaction. If we could challenge the user without sending a challenge to the user, we could authenticate in one HTTP transaction. And we can do exactly that by treating the URI as a challenge. * The first time, the user and service perform the authentication process. G Brown [Page 12] INTERNET-DRAFT RPA - Part Two 15 November 1996 * The user and service remember the session key (Kus), challenges (Cu and Cs), and timestamp (Ts). * When the user generates an HTTP request, he includes an Authorization header containing a response calculated as MD5(Kus + Z + Ns + Nu + Nr + Cs + Cu + Ts + method + URI + Kus) The method and URI are canonicalized by taking the big-endian Unicode representation and converting all characters to lowercase; the URI should not include the scheme://host:port. It always begins with a slash; for "http://www.foo.com" the one-character string "/" would be used. Now the authentication response is unique for each URI, and calculable only by the authenticated user, even without a unique challenge. This doesn't completely eliminate the risk of replay, of course, but an attacker can replay only a previously referenced URI during the window in which the service considers the session key to be valid. Is that acceptable? Sometimes. If we're reading Web pages, and the only impact of replay is that the attacker could re-read the page, it might be acceptable--after all, the attacker saw the page, anyway, when he captured it along with the original request. On the other hand, if we're charging the user per page, or if the request "did" something, replay might not be so harmless. One strategy is to maintain some history. In its simplest form, the service sets a flag for this session when it does something for which replay would be harmful. If the user tries reauthentication cheating, and the flag is set, the service forces reauthentication. Because the cheating response is based on Cu and Cs, and those values change during reauthentication, the correct response for a given URI changes after reauthentication. Thus, reauthentication creates a boundary after which previous requests cannot be replayed. Or the service can maintain a history of URIs for which replay would be harmful, and force reauthentication only if the user tries reauthentication cheating on one of those URIs. 5. Security Considerations This entire document is about security. G Brown [Page 13] INTERNET-DRAFT RPA - Part Two 15 November 1996 6. Author's Address Gary S. Brown CompuServe Incorporated 5000 Britton Rd P.O. Box 5000 Hilliard OH 43026-5000 USA +1 614 723 1127 This Internet-Draft expired on May 15, 1997. G Brown [Page 14]