Re: [TICTOC] [ntpwg] security ID submitted for review

["David L. Mills" <[e-mail]>]

Kurt,

Thanks for your reply.  As a window of history, the Autokey protocol was 
devised during several walks in Regents Park, London, in 1996.  I 
thought the most significant hazzard was the replay of client messages 
requiring extensive server cryptographic computations.  At that time, I 
did not think that a middle-man attack would be significant.  Both of 
those assumptions are probably inadvisable today. 

One problem with Autokey is the assumption that secret keys for each 
client can be manufactured using packet header fields and a hash of a 
server private value.  A middle-man can manipulate the header fields 
compromising the client private key.  There does not seem to be a 
solution for this problem other than requiring server state.

There are two problems with requiring a server state variable associated 
with each client.  The first is how to reliably match the server state 
with the client packet.  For public time servers with thousands of 
clients, the storage of something like a client certificate for each 
client could be prohibitive.  The storage could be dynamically allocated 
as a field in the MRU list maintained in the current NTP implimentation; 
however, the MRU list for several hundred entries has a lifetime of only 
15 seconds, at NIST, for example. 

In the current Autokey design, the MAC computation covers the header and 
all extension fields.  However, the key used for the MAC computation in 
packets with extension fields is not private.  Therefore, in a 
cut-and-paste attack, an intruder could mix and match extension fields 
and simply recalculate the MAC. 

RFC 5906 defines the extension field including the filestamp and 
timestamp.  The filestamp was designed to reliably associate the 
cryptographic media with the current version, as it can be frequently 
changed.  This might not be useful in current designs.  The timestamp 
field was designed to detect replays requiring expensive cryptographic 
computations which could take up to a second.  However, the "call gap" 
(minimum headway) provisions nicely deflect such attacks, so the 
timestamp field might not be useful in current designs.   Lastly, in the 
current design, each extension field is separately signed  at the time 
the data were constructed.  In current designs the signature might be 
computed at the time of transmission.

The more I think of it, the better I like a design where the existing 
MAC is replaced by a single extension field computed as the signature of 
the header.  I previously suggested this for broadcast mode, but I think 
this could be useful in other modes as well.  It would be hightly 
interesting to determine the overhead for such a signature using modern 
processors.  An experiment could be designed for the current NTP 
implementation.  It would consist of an adaptor routine that matches an 
OpenSSL signature routine to the authencrypt/authdecrypt interface 
routine.  I would do this myself, but with my current vision condition, 
it would not be practical.  For someone with knowledge of the NTP 
programming style, this would be an interesting weekend project.

In theis proposed design a busy server could be presented with thousands 
of signed client requests, each requiring  a signed server response.  
However, the server might ignore the client request signature on the 
expectation that the  server response could be reliably determined.  
Actually, the NTP on-wire protocol is very good for this.  As an aside, 
the protocol analysis and simulation described in the white paper at the 
NTP project page could be done for PTP as well.

Dave

Kurt Roeckx wrote:

>On Thu, Aug 01, 2013 at 08:42:08PM +0000, David L. Mills wrote:
>  
>
>>Guys,
>>
>>Here are some general comments on the ID.  I apologize if they are a
>>little rough as I have to dictate this message instead of my usual
>>polished wordsmithing.
>>
>>format.
>>The ID does not specify an extension field format.
>>    
>>
>
>The draft only seems to specify the the general idea of how it
>should work and doesn't seem to go into detail how this should
>work at the protocol level.  I assume he first wants to know if
>the solution is something that can work and that others see
>as a good solution before he works out the details.
>
>  
>
>>In particular, the private key for the message
>>authentication code (MAC) , if used, must be known beforehand.  This
>>is where the "cookie snatcher" issue comes up as described in
>>previous messages.  In particular, an important constraint for
>>normal operation is that extension fields are not used unless
>>required for cookie determination.
>>    
>>
>
>I'm not sure I understand your remark.  As far as I know the
>"cookie snatcher" you talk about is the normal "man in the middle"
>attack.  As I understand we would change the verification by
>replacing it by the normal X.509 verification as done for instance
>SSL/TLS.
>
>I assume you meant public key and note private key.
>
>  
>
>>Glaring vulnerability exists when using extension fields, as the key
>>value is public and must be known in advance.  This invites a
>>cut-and-paste attack, as well as header changes, even if
>>subsequently detected by the on-wire protocol.
>>    
>>
>
>I've seen this cut-and-paste attack mentioned before, but I've
>never really understood it.  As I understand things, the MAC
>is generated over the NTP + extensions.  I don't see how you
>can take the extension from 1 packet into an other and still
>have a valid MAC.  Or are you saying that you do something
>with a packet even when it doesn't have a valid MAC?
>
>  
>
>>certificate.
>>Both RFC 5906 and the ID require the server certificate.  The server
>>certificate can be conveyed, either off-line, or directly or
>>indirectly via a certificate trail as described in RFC 5906 or the
>>ID.  If obtained via the trail, there is a hazzard due to a
>>middle-man attack and masquerade.
>>    
>>
>
>I really don't understand how such a man in the middle attack
>can work.  Either the certificate + the chain validate, or they
>don't.  It is of course important to you check that the CommonName
>in the certificate matches the server you're trying to reach,
>and that the root CA is in your list of trusted CAs.  But there
>really isn't anything new or hard about this.
>
>  
>
>>In broadcast mode, the protocols described in RFC 5906 and the ID
>>are messy and with several possible vulnerabilities.  Here is a
>>suggestion that might be useful.  Define a new broadcast packet
>>format as the NTP header followed by one extension field
>>representing the MAC.  The extension field represents the signature
>>of the header computed using the private key of the server and can
>>be verified by the client using the public key in the server
>>certificate.  Originally, this was considered an unacceptable
>>dilution of accuracy due to the variability of the signature
>>computing time and the length of the signature.  However, at least
>>in modern NTP, the interleaved broadcast mode can be used to correct
>>for this overhead.  This assumes that the overhead for either the
>>server or client is reduced using fast modern processors.
>>    
>>
>
>I think that in all cases where it's no problem to keep track of
>all clients you should use the interleaved mode, and we can use
>something like the Diffie-Hellman key agreement for those.
>
>But there clearly is also a need for servers that can't keep
>the state of each client, and we need something else for that.
>So maybe we need to define 2 ways of working.
>
>
>
>Kurt
>
>  
>