Re: [TLS] Fully encrypted and authenticated headers (was Re: Encrypting record headers: practical for TLS 1.3 after all?)

[Jacob Appelbaum <[e-mail]>]

On 12/2/15, Eric Rescorla <[e-mail]> wrote:
> On Wed, Dec 2, 2015 at 1:07 AM, Bryan Ford <[e-mail]> wrote:
>
>> On 02 Dec 2015, at 06:02, Martin Thomson <[e-mail]>
>> wrote:
>> > On 1 December 2015 at 08:22, Bryan A Ford <[e-mail]>
>> > wrote:
>> >> The 2-byte length field in each record's header no longer indicates
>> >> the length of the *current* record but instead indicates the length of
>> >> the *next* record.
>> >
>> > Ensuring that you know the length of the *next* record is difficult
>> > and could dramatically degrade latency, or adding extra bogus padding
>> > or extra bogus records.  For instance, I can always send in bursts of
>> > two packets, a one octet packet that promises the remainder of the
>> > burst and one that promises a single octet packet.  At that point, I
>> > get to do what I've always done and you have gained little other than
>> > an increase in packet size of around 19 octets (best case).
>>
>> That type of inefficiency is extremely easy to avoid; please read the
>> rest
>> of my proposal where I discussed exactly that at length.  Yes, a
>> particularly stupid implementation could send everything in bursts of two
>> packets, but it’s ridiculously easy for a slightly smarter implementation
>> to avoid doing that.  And what you’ve gained is complete encryption and
>> integrity-checking of the whole TLS record before any part is
>> interpreted,
>> which seems like a nontrivial security improvement.
>
>
> It's not really clear to me what the anti-traffic-analysis benefit of your
> proposal
> is over and above just padding everything to a fixed size. That's certainly
> far
> easier for the implementation to do, especially in typical stacks where the
> application just calls SSL_Write (or whatever) and so there's no obvious
> API point to provide the "next length", so as a practical matter the stack
> will very often if not always be in "last block" mode.
>

I think that it eliminates all static distinguisher in the protocol
for all data covered by the encryption. That is a fantastically
wonderful benefit.

> The primary security improvement of your proposal seems to be that an
> active attacker can't generate a packet header that will put the TLS
> implementation in a deadlock state where it's waiting for more data that
> will never come. This seems of modest value in TLS [0] given that the
> attacker
> can cause the connection to be torn down by modifying any packet.
> I agree, that this is not exactly the same as leaving the connection
> deadlocked
> but it still effectively breaks the connection. In addition, I'm not an
> expert on TCP internals, but can't you also cause a similar deadlock by
> removing a TCP segment sent to the receiver and then ACKing it to the
> sender so that there is a gap in the TCP stream?

Yes, any TLS connection may be broken by TCP's total lack of
confidentiality, integrity or authenticity. It seems that normally
this happens at setup by IP:port blocking or during later transmission
of a selector/distinguisher that triggers an attack (TCP RST or
others). There is good work in this area by David Stainton with his
Honeybadger project - he actually classified and implemented most of
these TCP attacks to help detect QUANTUMINSERT attacks in the wild:

  https://www.noisebridge.net/pipermail/noisebridge-discuss/2015-April/046273.html

>
> -Ekr
>
> [0] This issue doesn't apply to DTLS because the stack will just move onto
> the next UDP datagram.

An off-path attacker can't do much with DTLS, if designed correctly.
Especially if they only see some packets - they'd only get the UDP
headers and then the rest should be uniformly distributed random data,
no? An attacker could thus only interfere during setup - which if
they're on path - we can't stop and expect; the PKI should (ha!) help
with this issue. After that point, especially with devices that move
or are otherwise multi-homed, would be able to spray encrypted packets
out to the network with little other than the UDP headers for an
attacker to use.

Bryan's proposal makes things strictly better with regard to a network
attacker - especially a partial view or off-path attacker who can only
inject packets or who require a selector to trigger an attack.
Effectively, an attacker will be forced to shut down the connection at
setup time or to use the layer separation issues with TCP at any
point. The key difference is that they will have the same information
as they did at setup time, which again seems as a strict improvement
over the current status quo.

All the best,
Jacob