[Babel-users] About an authentication extension

[Toke Høiland-Jørgensen]

Rodrigo Garcia <strysg at riseup.net> writes:

> El 08/09/17 a las 05:28, Toke Høiland-Jørgensen escribió:
>> Rodrigo Garcia <strysg at riseup.net> writes:
>> 
>>> Hello, i've been working on an extension to the babel protocol to
>>> authenticate updates using RSA. It's mostly an academic excercise and i
>>> have planned the first functional versions in a month or so.
>> 
>> What are you trying to achieve? What is your threat model?
>> 
>
> Hi, i wanted to reduce the risk of ip spoofing as an academic excercise.

I'm all for academic exercises, I'm just suggesting that it'll be
helpful to define (on the protocol level) what you are trying to protect
against. I.e., which nodes should be prevented from doing what?

> I tried 3 routers and two of them advertising the same network prefix,
> at first 1st and 2nd router exchange routing information and established
> a link then let them sending ICMP ping.
>
> Secondly, the 3rd router starts advertising the same prefix as the 2nd
> and, then the 3rd router is intetionally put closer than 2nd to make it
> advertise a better metric.
>
> After some time 1st router prefers to take the route to the 2nd
> network prefix through 3rd. 2nd stoped to receive the ICMP packets
> from 1st, and 3rd started receiving ICMP packets from 1st instead.

Right, I see. Are you familiar with the HMAC extension to babel
(RFC7298)? That does something different (it prevents nodes that don't
know the shared secret from participating in the network at all, but
does not restrict which prefixes each node can export). However, it may
be useful to read at least parts of it to help you formulate the
requirements for your own scheme.

>>> The main idea is to make that a *central entity* generates a set of
>>> encrypted tokens using a RSA *private key*. The set of tokens is
>>> unique to every node on the network and every node receives it's
>>> unique set of tokens, a network prefix and the *public key* from the
>>> central entity.
>> 
>> You're basically talking about signed tokens here, right? Why do they
>> need to be encrypted? This sounds a little bit like you're rolling your
>> own crypto (which is generally a bad idea). I'd suggest you make
>> absolutely sure you have your primitives defined, and also define the
>> problems you are going to solve (threat model etc).
>
> The tokens are not signed, they're just encrypted with a *private key*
> wich no single node owns. Instead a node must use the corresponding
> *public key* of the private key (wich is only used by a central entity
> to generate the encrypted set of tokens unique to every node)

But if everyone knows how to decrypt all the tokens they are not really
secret; so it basically becomes the same as a signature, no? Except if
it's *not* signed you may be able to spoof other values by changing the
ciphertext of a valid token you already own (not sure how susceptible
public crypto is to this)...

>>> Every decrypted token has the form:
>>>
>>> <network_prefix>_<randomnumber>
>>>
>>> example:
>>>
>>> 80.0.2.0/24_48003818
>>> 80.0.2.0/24_41106879
>> 
>> So what you are authenticating is that this router is authorised to
>> export this prefix? What's the purpose of the random number? How are you
>> preventing an intermediate router from just grabbing a token from
>> another router and reusing it?
>
> Yes, on an update a node must send a token which shows that it really
> owns the prefix it is being advertising.

So why not put that into a sub-TLV of the update?

> To prevent a node to just capture a token then reuse it a node should
> have a big ammount of encrypted tokens and send different ones, the
> random number is just to make then different and to make it more
> difficult to guess the private key.

So what happens if a node runs out of tokens? How is another node
supposed to deal with an update that has no corresponding token? Always
ignore it? For how long?

>>> I was planning to use encrypted tokens of 512 octets.
>> 
>> Holy cow; you realise that on some mediums a single token will take up
>> the entire available package size, right? And you presumably need a
>> token for each prefix? Why do you need such big signatures? ed25519
>> signatures fit in 64 bytes, for instance, and that will lower the
>> computational cost as well...
>> 
>
> Yes it is, in practical terms i was planning that only a set of
> important nodes use the authentication method.
>
>>> Every node must send a token from it's unique set of tokens on every
>>> route update. When a node receives an update, it tries to decrypt the
>>> token using the *public key* of the central entity and checks if the
>>> decrypted token contains the network prefix being advertised.
>> 
>> So the tokens are propagated across the entire network? That is *a lot*
>> of extra traffic.
>
> Yes it will be, but for flexibilty and network traffic sake, i think a
> node could forward also not authenticated updates (without the encrypted
> token) then if a node receives an update *with* the encrypted token it
> should *prefer* that update over common ones and if it is valid then
> forward it.

How does this interact with the other preference rules of Babel? Loop
avoidance in particular; if there's an non-feasible signed route but a
feasible unsigned route, is a node allowed to pick the unsigned one?

-Toke