Job, many thanks for this detailed feedback! I copy Nanog on my response below, since you sent there, but maybe we could move future emails off list to avoid spamming people not into this subject.
- It appears that BGP-iSec requires a deployment flag day per Autonomous
System, rather than doing security upgrades “one EBGP session at a
time” (a deployment model both RPKI-ROV and BGPsec support).
Thanks for this important comment; indeed, if something like BGP-iSec would be adopted as a standard, then such
flad day' requirement should indeed be avoided. One simple option would be to allow an adopting AS X to include in its relevant RPKI entry, a specific enforcement time’ T_E. Only after this time would another adopting AS, say AS Y, drop announcements with origin AS X but without a valid signature. AS X will select the enforcement time T_E to be sufficient in the future (upon deploying) to ensure that all of its routers will already be deployed at that time. But we don’t view BGP-iSec as a proposal for adoption; we expect few improvements in future work, in particular, improvements to its efficiency (which is currently comparable to, and even somewhat higher than, that of BGPsec, i.e., quite high). And additional evaluation is also very important (esp. from experts like you, Job - thanks!) and may lead to changes. For example, we may do an fix to the design soon to address an important comment from Joel Halpern (thanks Joel!).
- As noted in RFC 7132 , Google’s , Fastly’s  response to
the FCC “inquiry into internet routing vulnerabilities” which your
paper cites; BGPsec was consciously not designed to address route
leaks, as leaks are not precluded by the specification for BGP and
might be the result of a local policy that is not publicly disclosed.
That’s correct. But, again, BGP-iSec is an academic paper, not a proposed standard. In practically all papers on BGP security, we adopt the valley-free routing model which basically forbids leaks. If this is modified to a standard, I’m sure this issue can be addressed, similarly to how it is dealt with in the OTC RFC.
To me the sentence “Even under full adoption, BGPsec does not prevent
route leaks” reads contentious, because of an implicit assertion that
BGPsec was intended to address route leaks.
Good point, we should clarify this indeed, thanks!
To me the above reads as “Even under full adoption of IPv6, world
hunger will not be achieved”. E.g. seems a false equivalence is drawn.
I admit that sometimes it appears to me that full adoption of IPv6 will indeed occur only after we solve world hunger.
- It appears BGP-iSec took some inspiration from ASPA, but instead of
signalling the list of providers out-of-band (which ASPA does via the
RPKI publication system); BGP-iSec proposed an in-band signal in the
form of ‘UP’ messages encapsulated inside BGP Path Attributes.
Actually, I think we became aware of ASPA only after that part of the design
this suggests that BGP-iSec requires a flag day for all EBGP routers
in a given Autonomous System;
Here you’re wrong. There is no problem if different EBGP routers adopt the UP mechanism at different times; the mechanism will be implemented only for announcements which passed via an adopting router, that’s all, no harm.
- The bold assertion made by the BGP-iSec authors that BGPSec offers
“meagre benefits in partial adoption” to me seems a subjective one.
You’re right, we should clarify that we merely referred to the reductions in hijack rates in simulations of the defense (under different adoption percentages). Our expression was too extreme and may give wrong impression; thanks for pointing this out.
Consider when two BGPSec-capable networks interconnect, both parties
immediately benefit from having secured that interconnection point.
For example, if this happens to be the interconnection between two
major cloud providers, the advantages are massive and can positively
benefit billions of indirect constituents.
Let’s say that these large cloud ASes, say AS A an AS Z have a direct connection (BGP peering) and both adopt BGPsec. What benefit do you see (to billions of entities)? In particular, do you think this prevents a attacker from hijacking traffic sent from A to Z? I’m not sure. Consider AS 666 who is a customer of AS A, and let 1.2.3/24 be a prefix of cloud/AS Z. Now, AS 666 may send an announcement with origin Z and path 666-Z to AS A; this may be unsigned (without even receiving announcement of 1.2.3/24 from Z) or even signed (if Z received the announcement from Z - e.g., when 666 is a customer of Z). I think that in both cases, A ends up sending the traffic via 666. So, can you clarify the benefits? I may miss something.
Similarly, in the past I’ve
argued that only the biggest networks in the world need to do RPKI-ROV
for all of the Internet to take benefit from that deployment action by
a single small group. Comparing RPKI-ROV & BGPSec is apples and
I definitely agree with this last statement. RPKI-ROV does not automatically downgrade once an announcement exits an island of deploying ASes.
but the point stands that in an Internet with increased
centralization we have to rethink what exactly the consequences are of
so-called ‘partial deployments’ of security features.
But that’s exactly what we try to do by using simulations to measure the impact.
I enjoyed reading the BGP-iSec paper and certainly view and appreciate
it as an interesting perspective on the routing security problem space;
but I think for very good reasons AS_PATH integrity (BGPsec) and AS_PATH
route leaks (OPEN Policy, OTC, ASPA) are addressed as independent
technology extensions to the routing stack.
I’m not arguing against separation. OTOH, our results show that OTC becomes effective even against intentional attack, if protected by the same integrity mechanisms we use to protect the AS-path.
Sometimes, trying to kill two or three birds with one stone
inadvertently introduces significant cost in unexpected places,
presenting surprising barriers to deployment.
True. And deployment and standardization are very important and challenging. BGP-iSec, at this point, is just an academic study studying some new ideas and evaluating their impact in specific configurations, under specific assumptions etc.; hopefully, this may provide some help to the community in improving BGP security.