Here's a proposal Nick Mathewson and I just wrote for ticket #11157.
--- Begin proposal body --- Filename: xxx-consensus-hash-chaining.txt Title: Consensus Hash Chaining Author: Nick Mathewson, Andrea Shepard Created: 06-Jan-2015 Status: Draft
1. Introduction and overview
To avoid some categories of attacks against directory authorities and their keys, it would be handy to have an explicit hash chain in consensuses.
2. Directory authority operation
We add the following field to votes and consensuses:
previous-consensus ISOTIME [SP HashName "=" Base16]* NL
where HashName is any keyword.
This field may occur any number of times.
The date in a previous-consensus line in a vote is the valid-after date of the consensus the line refers to. The hash should be computed over the signed portion of the consensus document. A directory authority should include a previous-consensus line for a consensus using all hashes it supports for all consensuses it knows which are still valid, together with the two most recently expired ones.
When this proposal is implemented, a new consensus method should be allocated for adding previous-consensus lines to the consensus.
A previous-consensus line is included in the consensus if and only if a line with that date was listed by more than half of the authorities whose votes are under consideration. A hash is included in that line if the hash was listed by more than half of the authorities whose votes are under consideration. Hashes are sorted lexically with a line by hashname; dates are sorted in temporal order.
If, when computing a consensus, the authorities find that any previous-consensus line is *incompatible* with another, they must issue a loud warning. Two lines are incompatible if they have the same ISOTIME, but different values for the the same HashName.
The hash "sha256" is mandatory.
3. Client and cache operation
All parties receiving consensus documents should validate previous-consensus lines, and complain loudly if a hash fails to match.
When a party receives a consensus document, it SHOULD check all previous-consensus lines against any previous consensuses it has retained, and if a hash fails to match it SHOULD warn loudly in the log mentioning the specific hashes and valid-after times in question, and store both the new consensus containing the mismatching hashes and the old consensus being checked for later analysis. An option SHOULD be provided to disable operation as a client or as a hidden service if this occurs.
All relying parties SHOULD by default retain all valid consensuses they download plus two; but see "Security considerations" below.
If a hash is not mismatched, the relying party may nonetheless be unable to validate the chain: either because there is a gap in the chain itself, or because the relying party does not have any of the consensuses that the latest consensus mentions. If this happens, the relying party should log a warning stating the specific cause, the hashes and valid-after time of both the consensus containing the unverifiable previous-consensus line and the hashes and valid-after time of the line for each such line, and retain a copy of the consensus document in question. A relying party MAY provide an option to disable operation as a client or hidden service in this event, but due to the risk that breaks in the chain may occur accidentally, such an option SHOULD be disabled by default if provided.
If a relying party starts up and finds only very old consensuses such that no previous-consensus lines can be verified, it should log a notice of the gap along the lines of "consensus (date, hash) is quite new. Can't chain back to old consensus (date, hash)". If it has no old consensuses at all, it should log an info-level message of the form "we got consensus (date, hash). We haven't got any older consensuses, so we won't do any hash chain verification"
4. Security Considerations:
* Retaining consensus documents on clients might leak information about when the client was active if a disk is later stolen or the client compromised. This should be documented somewhere and an option to disable (but thereby also disable verifying previous-consensus hashes) should be provided.
* Clients MAY offer the option to retain previous consensuses in memory only to allow for validation without the potential disk leak. --- End proposal body ---