Salted Challenge Response Authentication Mechanism

In cryptography, the Salted Challenge Response Authentication Mechanism (SCRAM) is a family of modern, password-based challenge–response authentication mechanisms providing authentication of a user to a server. As it is specified for Simple Authentication and Security Layer (SASL), it can be used for password-based logins to services like SMTP and IMAP (e-mail), or XMPP (chat). For XMPP, supporting it is mandatory.^[1]

Motivation

Alice wants to log into Bob's server. She needs to prove she is the one she pretends to be. For solving this authentication problem, Alice and Bob have agreed upon a password, which Alice knows, and which Bob knows how to verify. Now Alice could send her password over an unencrypted connection to Bob in a clear text form, for him to verify. That would however make the password accessible to Mallory, who is wiretapping the line. Alice and Bob could try to bypass this by encrypting the connection. However, Alice doesn't know whether the encryption was set up by Bob, and not by Mallory by doing a man-in-the-middle attack. Therefore, Alice sends a hashed version of her password instead, like in CRAM-MD5 or DIGEST-MD5. As it is a hash, Mallory doesn't get the password itself. And because the hash is salted with a challenge, Mallory could use it only for one login process. However, Alice wants to give some confidential information to Bob, and she wants to be sure it's Bob and not Mallory.

For solving this, Bob has registered himself to a Certificate Authority (CA), which signed his certificate. Alice could solely rely on that signature system, but she knows it has weaknesses. To give her additional assurance that there is no man-in-the-middle attack, Bob creates a proof that he knows the password (or a salted hash thereof), and includes his certificate into this proof. This inclusion is called channel binding, as the lower encryption channel is 'bound' to the higher application channel.

Alice then has an authentication of Bob, and Bob has authentication of Alice. Taken together, they have mutual authentication. DIGEST-MD5 already contained an attempt to achieve mutual authentication, but it has often been implemented in a wrong way.^[2]

When Mallory runs a man in the middle attack and forges a CA signature, he could retrieve a hash of the password. But he couldn't impersonate Alice even for a single login session, as Alice included into her hash the encryption key of Mallory, resulting in a login-fail from Bob. To make a fully transparent attack, Mallory would need to know the password used by Alice, or the secret encryption key of Bob.

Bob has heard of data breaches of server databases, and he decided that he doesn't want to store the passwords of his users in clear text. He has heard of the CRAM-MD5 and DIGEST-MD5 login schemes, but he knows, for offering these login schemes to his users, he would have to store weakly hashed, un-salted passwords. He doesn't like the idea, and therefore he chooses to demand the passwords in plain text. Then he can hash them with secure hashing schemes like bcrypt, scrypt or PBKDF2, and salt them as he wants. However, then Bob and Alice would still face the problems described above. To solve this problem, they use SCRAM, where Bob can store his password in a salted format, using PBKDF2. During login, Bob sends Alice his salt and the iteration count of the PBKDF2 algorithm, and then Alice uses these to calculate the hashed password that Bob has in his database. All further calculations in SCRAM base on this value which both know.

Protocol overview

Although all clients and servers have to support the SHA-1 hashing algorithm, SCRAM is, unlike CRAM-MD5 or DIGEST-MD5, independent from the underlying hash function.^[3] All hash functions defined by the IANA can be used instead. As mentioned in the Motivation section, SCRAM uses the PBKDF2 mechanism, which increases the strength against brute-force attacks, when a data leak has happened on the server. Let be the selected hash function, given by the name of the algorithm advertised by the server and chosen by the client. 'SCRAM-SHA1' for instance, uses SHA1 as hash function.

Messages

RFC 5802 names four consecutive messages between server and client:

client-first

The client-first message consists of a gs2-cbind-flag, the desired , and a randomly generated client nonce .

server-first

The server appends to this client nonce his own nonce , and adds it to the server-first message, which also contains a used by the server for salting the user's password hash, and an iteration count indicator .

client-final

After that the client sends the client-final message, which contains c-bind-input, the concatenation of the client and the server nonce, and a proof of all the messages sent, and the contents of client-final up to the proof.

server-final 

The communication closes with the server-final message, which contains the server proof .

Stored password

The stored password is calculated as follows:

,

where is defined as PBKDF2 (HMAC, , , , output length of ).

For every user, the server only has to store the username and together with and , but not the clear text password itself.

Proofs

The client and the server prove to each other they have the same variable, consisting of:

The proofs are calculated as follows:

where defines the XOR operation. and are verbatim strings.

Channel binding

The term channel binding describes the man-in-the-middle attack prevention strategy to 'bind' an application layer, which provides mutual authentication, to a lower (mostly encryption) layer, ensuring that the endpoints of a connection are the same in both layers. There are two general directions for channel binding: unique and endpoint channel binding. The first ensures that a specific connection is used, the second that the endpoints are the same.

There are several channel binding types, where every single type has a channel binding unique prefix.^[4] Every channel binding type specifies the content of the channel binding data, which provides unique information over the channel and the endpoints. For instance, for the tls-server-end-point channel binding, it is the server's TLS certificate.^[5] An example use case of channel binding with SCRAM as application layer, could be with Transport Layer Security (TLS) as lower layer. While TLS protects from passive eavesdropping, it alone doesn't prevent man-in-the-middle attacks. For this, the endpoints need to assure their identities to each other, which is provided by SCRAM.

The gs2-cbind-flag SCRAM variable specifies whether the client supports channel binding or not, or thinks the server doesn't support channel binding, and c-bind-input contains the gs2-cbind-flag together with the channel binding unique prefix and the channel binding data themselves.

Channel binding is optional in SCRAM, and the gs2-cbind-flag variable prevents from downgrade attacks.

When a server supports channel binding, it adds the character sequence '-PLUS' to the advertised SCRAM algorithm name.

Strengths

• Strong password storage: When implemented in a right way, the server can store the passwords in a salted, iterated hash format, making offline attacks harder, and decreasing the impact of database breaches.^[6]
• Simplicity: Implementing SCRAM is easier^[7] than DIGEST-MD5.^[8]
• International interoperability: the RFC requires UTF-8 to be used for usernames and passwords, unlike CRAM-MD5.^[7][9]
• Because only the salted and hashed version of a password is used in the whole login process, and the salt on the server doesn't change, a client storing passwords can store the hashed versions, and not expose the clear text password to attackers. Such hashed versions are bound to one server, which makes this useful on password reuse.^[10]

References

External links

• RFC 5802, SCRAM for SASL and GSS-API
• SCRAM in HTTP (currently in draft state)
• GNU Network Security Labyrinth (presentation similar to Motivation section)