ARM4FS Designfoo: Difference between revisions

From
Jump to navigation Jump to search
 
(10 intermediate revisions by the same user not shown)
Line 35: Line 35:
<dt>ObjectKey</dt> <dd>URI, preferably URL with (schema specific) embedded hash of the contents. For starters: <pre>http://SP/get?h=hash</pre>.</dd>
<dt>ObjectKey</dt> <dd>URI, preferably URL with (schema specific) embedded hash of the contents. For starters: <pre>http://SP/get?h=hash</pre>.</dd>
</dl>
</dl>

== Musings On The Server<->Client Protocol ==

The Reputation Provider acts as a server and listens on port 2323.

=== Encryption ===
The server<->client protocol, or the sessions, should provide protection against:
* all passive attacks
* most active attacks
Specifically it seems to only be possible to protect against active attacks (man in the middle) if a trust relationship has already been established, e.g. the client knows the server's public key. (In order to provide confidentiality the server does not need to know the client's public key. This key only becomes important for authenticating clients. The user creation process is possible even for an unauthenticated client.)

To minimize the possibility of protocol design or implementation errors it would be best to use an already existing protocol and implementation, e.g. SSL/TLS and OpenSSL. Problem: SSL uses X.509 certificates and not raw keys. Certificates pose some dangers of introducing accidental side channels, for example through serial numbers, but add next to no additional value. Also: SSL is quite a complex protocol and widely deployed implementation errors (ASN.1!) have not been unheard of.

A homemade protocol would look something like this:
# diffie-hellmann between client and server
# server signs own diffie-hellmann parameters with own secret key
# when the client has verified this signature it knows that it is in fact talking to the right server
# connection is then in a state in which the client may
#* transmit a user creation request, '''or'''
#* authenticate itself by sending a signed version of its own diffie-hellmann parameter

==== Discussion ====

{| cellpadding="5" cellspacing="0" style="border-collapse: collapse;"
!
! style="border: 1px solid black;" | own protocol
! style="border: 1px solid black;" | SSL/TLS
|-
! style="border: 1px solid black;" | pro
| style="border: 1px solid black; vertical-align: top;" | <ul><li>uses raw keys, limited side channel possibilities</li></ul>
| style="border: 1px solid black; vertical-align: top;" | <ul><li>tried and proven design</li><li>various interoperable implementations</li><li>may be possible to verify server identity through regular PKI</li></ul>
|-
! style="border: 1px solid black;" | contra
| style="border: 1px solid black; vertical-align: top;" | <ul><li>not thoroughly analysed, may be weak</li></ul>
| style="border: 1px solid black; vertical-align: top;" | <ul><li>some degree of bloat (ASN.1, X.509)</li></ul>
|}

One compromise approach might be: Use SSL and a 'real' certificate for the server (allows verification through PKI) but only a raw key for the client. Then when the connection is set up and authentication is necessary, have the client transmit its public key and sign something with that key. Preferably some secret from the SSL connection (session key or master key perhaps?), but it's probably not possible to get access to this. Alternatively let the server send a challenge and have the client sign that (be wary of replay attacks and the such).

=== Payload ===
The cleartext protocol is simple and text-based. All commands are sent by the client. All commands and responses are single lines that end with a single newline ("\n"). The terminating newline is implied and not mentioned explicitly for the rest of this document because it is not technically considered to be a part of the command/response. A command or response MUST NOT have any newline anywhere except for the terminating newline.

Failure to follow any of the regulations in this protocol MAY lead to immediate termination of the connection.

==== Error Responses ====
Error responses start with an unhappy smiley, a single space and a three-digit error code, followed by an optional human readable error message, followed by an optional parameter section that starts with a semi-colon. They MUST thus match the following regular expression (PCRE syntax):
/^:-\( +\d{3}[^;]*(?:;.*)?$/

'''Error codes'''
{|
! Code !! Message !! Explanation
|-
! 400
| Bad command || The command was malformed or invalid. Note that the server has the option to immediately terminate the connection when it encounters a protocol error and is not obliged to send this response and keep going.
|-
! 401
| Authorization required || The command was not executed because the client must authenticate first, see the section on Authentication.
|}

==== Positive Responses ====
Positive responses start with a happy smiley, a single space and command specific data. Thus they MUST match the following regular expression (PCRE syntax):
/^:-\) .*$/

==== Authentication ====

Connections can be in two states: authenticated and unauthenticated. On unauthenticated connections only a minor subset of commands is available. When an unauthenticated client tries to access a command that needs authentication the server either sends a 401 error code or terminates the connection.

In order to initiate the authentication procedure:
# The client sends an <code>challenge</code> command with no parameter.
# The server sends a positive response with the challenge.
# The client sends an <code>authenticate</code> command with its own public key and the signed challenge
# Either
#* the server acknowledges the successful authentication with a positive response, if the signature checks out, or
#* should the authentication not be successful for any reason, then the server silently terminates the connection with no error message.

Authentication may be tried exactly once per connection. No other command may be sent between the start of the procedure (the <code>challenge</code> command) and the reception of the positive response to the <code>authenticate</code> command. Any deviation from this protocol is forbidden and MUST lead to a termination of the connection.

Latest revision as of 01:57, 13 May 2006


UserID
Public/private key pair, maybe with additional attributes
User management
  • RP has a database with two unrelated tables:
    • UserIDs and accompanying data
    • Realnames/IDs (X.509-DN?) who already have an account.
  • User creation, in cooperation with an identity provider:
    1. Identity provider asserts identity
    2. Reputation provider checks whether this identity already has an account, refusing creation if this is the case (counters Sybil attack)
    3. Create a new UserID and enter it into the database; at the same time enter the realname as 'used' into the database.
    After this procedure there is no further correlation between realname and UserID!
  • Two possible approaches to account for eventual deletion of UserIDs together with their realname entries (makes it possible to apply for another UserID thereafter):
    • Save a hash of the realname together with the UserID. Bad idea, because a simple scan could correlate the UserID and realname.
    • Give a signed (+ encrypted?) realname/UserID pair to the user on account creation. The user can then either delete this pair if she is positive that she won't need it in the future (or is afraid of the risk that this pair may be found on her computer), or use it in the future to have the reputation provider delete both entries (UserID table and realname table).
Document submission
session = RP.createSession(UserID) ObjectKey = SP.storeTemporary(Content, timeout) AuthorTag = session.createAuthorTag(ObjectKey) SP.storePermanent(AuthorTag) session.destroy()
Reputation query
session = RP.createSession(UserID) session.getReputation(AuthorTag) session.destroy()
Content feedback
session = RP.createSession(UserID) session.submitFeedback(AuthorTag, feedback) session.destroy()
AuthorTag
signed_RP(UserPseudonym, ObjectKey)
Sessions
Authenticated, encrypted channel to the Reputation Provider. Maybe integrate Off-the-Record here? Would that be beneficial? In any case: it's fine to route the channel through TOR or similar.
createAuthorTag()
UserPseudonym_x is a completely random value which will only be used to lookup the UserID in the Reputation Provider's database. Rationale: If anonymity is supposed to be the highest goal of this system then it won't be such a good idea to have any relationship between the public UserPseudonym and the secret UserID. Even encrypting isn't a good idea: If the encryption key is broken or compromised then all anonymity for all users for all past transactions would be void. (Break one, break all.) It might be possible to segment the effects of possible key compromises by using multiple keys (either separated by time or by user group), but that's only mildly effective and has problems of its own (e.g. by correlating on the key that was used it would be possible for an attacker to find out when an AuthorTag has been created, or which group a user belongs to; now that I think about it: this could be countered by randomly using many keys, but still doesn't look very good). The underlying threat model is that an administrator of the system might be 'convinced' into smuggling the key into the hands of the attacker. Certainly it would be a lot harder to transmit a complete database of UserID<->UserPseudonym relations than it would be to transmit a single key (about 4096 bits). On the other hand: It is easier to safeguard a key (e.g. in a security module such as a smart card) than it is to safeguard a complete database. Maybe we should combine both approaches? The public UserPseudonym would then be an encrypted random string. Note: The signature key for AuthorTag signatures is not so much of a problem. If this key is compromised then only the correctness of the system will be endangered and not the anonymity of its users.
ObjectKey
URI, preferably URL with (schema specific) embedded hash of the contents. For starters: 
http://SP/get?h=hash
.
Retrieved from "https://sarwiki.informatik.hu-berlin.de/index.php?title=ARM4FS_Designfoo&oldid=5593"