Filename: 191-mitm-bridge-detection-resistance.txt
Title: Bridge Detection Resistance against MITM-capable Adversaries
Author: George Kadianakis
Created: 07 Nov 2011
Status: Obsolete
1. Overview
Proposals 187, 189 and 190 make the first steps toward scanning
resistant bridges. They attempt to block attacks from censoring
adversaries who provoke bridges into speaking the Tor protocol.
An attack vector that hasn't been explored in those previous
proposals is that of an adversary capable of performing Man In The
Middle attacks to Tor clients. At the moment, Tor clients using the
v3 link protocol have no way to detect such an MITM attack, and
will gladly send a VERSIONS or AUTHORIZE cell to the MITMed
connection, thereby revealing the Tor protocol and thus the bridge.
This proposal introduces a way for clients to detect an MITMed SSL
connection, allowing them to protect against the above attack.
2. Motivation
When the v3 link handshake protocol is performed, Tor's SSL
handshake is performed with the server sending a self-signed
certificate and the client blindly accepting it. This allows the
adversary to perform an MITM attack.
A Tor client must detect the MITM attack before he initiates the
Tor protocol by sending a VERSIONS or AUTHORIZE cell. A good
moment to detect such an MITM attack is during the SSL handshake.
To achieve that, bridge operators provide their bridge users with a
hash digest of the public-key certificate their bridge is using for
SSL. Bridge clients store that hash digest locally and associate it
with that specific bridge. Bridge clients who have "pinned" a
bridge to a certificate "fingerprint" can thereafter validate that
their SSL connection peer is the intended bridge.
Of course, the hash digest must be provided to users out-of-band
and before the actual SSL handshake. Usually, the bridge operator
gives the hash digest to her bridge users along with the rest of
the bridge credentials, like the bridge's address and port.
3. Security implications
Bridge clients who have pinned a bridge to a certificate
fingerprint will be able to detect an MITMing adversary in time.
If after detection they act as an innocuous Internet
client, they can successfully remove suspicion from the SSL
connection and subvert bridge detection.
Pinning a certificate fingerprint and detecting an MITMing attacker
does not automatically alleviate suspicions from the bridge or the
client. Clients must have a behavior to follow after detecting the
MITM attack so that they look like innocent Netizens. This proposal
does not try to specify such a behavior.
Implementation and use of this scheme does not render bridges and
clients immune to scanning or DPI attacks. This scheme should be
used along with bridge client authorization schemes like the ones
detailed in proposal 190.
4. Tor Implementation
4.1. Certificate fingerprint creation
The certificate fingerprints used on this scheme MUST be computed
by applying the SHA256 cryptographic hash function upon the ASN.1
DER encoding of a public-key certificate, then truncating the hash
output to 12 bytes, encoding it to RFC4648 Base32 and omitting any
trailing padding '='.
4.2. Bridge side implementation
Tor bridge implementations SHOULD provide a command line option
that exports a fully equipped Bridge line containing the bridge
address and port, the link certificate fingerprint, and any other
enabled Bridge options, so that bridge operators can easily send it
to their users.
In the case of expiring SSL certificates, Tor bridge
implementations SHOULD warn the bridge operator a sensible amount
of time before the expiration, so that she can warn her clients and
potentially rotate the certificate herself.
4.3. Client side implementation
Tor client implementations MUST extend their Bridge line format to
support bridge SSL certificate fingerprints. The new format is:
Bridge <method> <address:port> [["keyid="]<id-fingerprint>] \
["shared_secret="<shared_secret>] ["link_cert_fpr="<fingerprint>]
where <fingerprint> is the bridge's SSL certificate fingerprint.
Tor clients who use bridges and want to pin their SSL certificates
must specify the bridge's SSL certificate fingerprint as in:
Bridge 12.34.56.78 shared_secret=934caff420aa7852b855 \
link_cert_fpr=GM4GEMBXGEZGKOJQMJSWINZSHFSGMOBRMYZGCMQ
4.4. Implementation prerequisites
Tor bridges currently rotate their SSL certificates every 2
hours. This not only acts as a fingerprint for the bridges, but it
also acts as a blocker for this proposal.
Tor trac ticket #4390 and proposal YYY were created to resolve this
issue.
5. Other ideas
5.1. Certificate tagging using a shared secret
Another idea worth considering is having the bridge use the shared
secret from proposal 190 to embed a "secret message" on her
certificate, which could only be understood by a client who knows
that shared secret, essentially authenticating the bridge.
Specifically, the bridge would "tag" the Serial Number (or any
other covert field) of her certificate with the (potentially
truncated) HMAC of her link public key, using the shared secret of
proposal 190 as the key: HMAC(shared_secret, link_public_key).
A client knowing the shared secret would be able to verify the
'link_public_key' and authenticate the bridge, and since the Serial
Number field is usually composed of random bytes a probing attacker
would not notice the "tagging" of the certificate.
Arguments for this scheme are that it:
a) doesn't need extra bridge credentials apart from the shared secret
of prop190.
b) doesn't need any maintenance in case of certificate expiration.
Arguments against this scheme are:
a) In the case of self-signed certificates, OpenSSL creates an
8-bytes random Serial number, and we would probably need
something more than 8-bytes to tag. There are not many other
covert fields in SSL certificates mutable by vanilla OpenSSL.
b) It complicates the scheme, and if not implemented and researched
wisely it might also make it fingerprintable.
c) We most probably won't be able to tag CA-signed certificates.
6. Discussion
6.1. In section 4.1, why do you truncate the SHA256 output to 12 bytes?!
Bridge credentials are frequently propagated by word of mouth or
are physically written down, which renders the occult Base64
encoding unsatisfactory. The 104 characters Base32 encoding or the
64 characters hex representation of the SHA256 output would also be
too much bloat.
By truncating the SHA256 output to 12 bytes and encoding it with
Base32, we get 39 characters of readable and easy to transcribe
output, and sufficient security. Finally, dividing '39' by the
golden ratio gives us about 24.10!
7. Acknowledgements
Thanks to Robert Ransom for his great help and suggestions on
devising this scheme and writing this proposal!