Re: [TLS] Encrypting record headers: practical for TLS 1.3 after all?

[Bryan A Ford <[e-mail]>]

Thanks Bill for the feedback.

On 11/29/15 6:11 PM, Bill Cox wrote:
> Thanks for the detailed description of why we might want to obfuscate
> TLS record lengths.  From a security point of view, the only potential
> attack vector this might create is that the attacker will know in many
> cases the plain text of the first 5 bytes.  A weakness in the stream
> cipher might be more easily exploited in this case.  IIUC, we feel
> pretty good about the latest AEAD ciphers, and guessing the clear text
> of many bytes in the record already is not too hard.  Is that right? 
> So, from my fairly uninformed point of view, this seems like a
> reasonable upgrade.

Just a nitpicky semantic correction: my proposal doesn't "create" the
attack vector you describe (that the attacker will know the plaintext of
the 5-byte header).  That weakness is already there in TLS 1.3 as
currently specified, because the TLS header is unencrypted so the
attacker will *definitely* know the 5-byte header. :)  This is merely an
existing weakness that is not fixed by my proposal.

> I have one concern.  These same boxes that hide record lengths by
> merging and breaking up packets arbitrarily likely would deliver the
> packets not well aligned to record boundaries.  Is this already the
> case?  

It's already the case that middleboxes of various kinds re-segment
(merge and break up along different boundaries) TCP streams for various
reasons, often as an unintentional side-effect of interposing on the
stream by "accepting" the TCP connection (pretending to be an endpoint)
on one of the middlebox's interfaces and opening a new outgoing
connection on the other interface, and just forwarding the traffic
between them. In fact, Performance Enhancing Proxies (PEPs) are one
class of middlebox that often do this intentionally, so they can get
better control over and fiddle with the way congestion control works
over "troublesome" types of links like high-bandwidth-delay-product
(e.g., satellite) links.

> If these boxes currently inspect packets when breaking them up to
> split packets along record boundaries, encrypting the lengths will
> defeat this efficiency hack.  We might cause an increase in network
> latency in this case.  Do any of the routers out there currently use the
> record length field when splitting up packets?

If I understand your question correctly, you're asking if encrypting TLS
record lengths might break middleboxes that currently want to track TLS
records for some reason, e.g., perhaps to re-segment the TCP stream to
correspond to TLS records (which might "fix" or "revert" the effect of
another middlebox in the path that re-segments TCP in a TLS-oblivious
fashion).  I don't know of any middleboxes that do precisely this, but
it's not inconceivable that they could exist.

More generally, since basically all manner of evil/broken middleboxes do
exist, it's entirely possible that most any change to TLS's wire
encoding *could* disagree with some middleboxes.  That's not specific to
the change I'm proposing; any wire-visible change could trigger unknown
middlebox badness.  The only solution to this I'm aware of is empirical:
try the change, deploy it experimentally at first, and see what the pain
level is, i.e., see if we find any middleboxes that complain and if so
how prevalent they are.  But since TLS 1.3 is already changing a lot of
other prominently wire-visible stuff with respect to prior versions of
TLS that might equally well break middleboxes, it's not clear that my
proposed change in particular is more likely to cause pain than any of
the other changes already made, and in general I don't think we should
let a few broken/evil middleboxes prevent meaningful evolution of the
TLS protocol.

Cheers
Bryan

> 
> Bill