"general badness" AS-based reputation system
mkarir at merit.edu
Mon Sep 26 00:11:50 CDT 2011
On Sep 25, 2011, at 11:31 PM, Tom Vest wrote:
> On Sep 25, 2011, at 9:23 PM, Manish Karir wrote:
>> On Sep 25, 2011, at 6:31 PM, nanog-request at nanog.org wrote:
>>> Message: 9
>>> Date: Sun, 25 Sep 2011 18:37:17 +0300
>>> From: Gadi Evron <ge at linuxbox.org>
>>> To: nanog at nanog.org
>>> Subject: "general badness" AS-based reputation system
>>> Message-ID: <4E7F4AAD.8020400 at linuxbox.org>
>>> Content-Type: text/plain; charset=ISO-8859-1; format=flowed
>>> Having run one of these in the past, when take-downs of C&Cs was still
>>> semi-useful, my ethos on this is problematic, however, I am as of yet
>>> undecided as to this one. An AS-based reputation system for all sorts of
>>> In my opinion, third-party security based AS-reputation systems will
>>> eventually become de-facto border filtering systems for ISPs, but that
>>> day is still not here, as that is still socially unacceptable in our
>>> circles, and will remain so until it becomes _necessary_.
>>> Regardless of my musings of Operators World cultural future, this
>>> systems seems rather interesting, and no doubt you'd want to take a look
>>> at your listing.
>> We tried to outline some of the challenges of building such a system in our NANOG52 presentation:
>> In particular see slide 4. where we tried to lay down what we think the requirements are for a socially acceptable
>> reputation system.
>> With a bit of luck we might be able to announce the release of our system before the next NANOG mtg, but in
>> my opinion collating host reputation reports is just a small and the easiest part of the effort. The key is in
>> solving the challenges of allowing (and incentivizing) participation and being robust to false information
>> Comments are welcome.
> Hi Manish,
> Looks like very interesting work.
> Does the system that you'll be announcing provide some means of coming to terms with challenges like the following?
> 1. Many large operators administer multiple ASNs, but some of the resulting AS sibling relationships may not be identifiable as such based on public-facing whois records, or interconnection relationships, or any other public data sources. Does your system incorporate some means of attributing reputation-related information at the (multi-AS) institutional level -- even when the contours of such institutions are not self-evident?
> 2. Some members of the ARIN community have recently floated a policy proposal which if approved would make ASNs transferable, and some supporters of that proposal have argued that RIR involvement in such transfers should be strictly limited to the passive recording of whatever information is voluntarily disclosed by the transacting parties, if and when a disclosure is made. Does your system ascribe reputation "strictly" to specific ASNs, such that declared changes in an ASN's "ownership"/control would not affect that ASNs accumulated reputation record to-date? Alternately, if declared changes to an ASN's ownership would affect (e.g., restart) an ASN's reputation history, will your system incorporate some mechanism for assessing the material/operational "substance" of ASN re-registration events (e.g., to filter for possible "re-registrations of convenience")?
> I like to ask these sort of questions (for any/all proposed systems like this) because it seems to me that any system that associates increasing value with a cumulative record of consistent "approved" behavior over time must take extra care not to provide *asymmetrical* opportunities (i.e., to some but not all participants) to isolate and sanitize their own "disapproved" behavior, thereby leaving their longstanding (favorable) reputations intact.
> Note that this problem is *not* reducible to the idea that a reputation system must be absolutely infallible. Obviously it is not reasonable to demand something that is impossible to deliver. However, it is altogether reasonable to demand that any system that is intentionally designed to produce a new, endogenously-driven reputation-based hierarchy of operational entities does something more than just recreate and reinforce pre-existing hierarchies that are completely orthogonal to "reputation."
> Look forward to hearing more!
At what granularity reputation is useful is an excellent question. Obviously we already have lots of single data source based host reputation sources.
Other possible aggregations are prefixes, ASNs, and as you suggest organizations (which might be multi-ASN). Another extreme possible aggregation is country.
In my opinion BGP prefix is the right level of aggregation up to be actually useful rather than narrow host reputation lists. You might expect hosts in a
prefix to be under the same security policy regime and hence have similar
likelihood of future malicious behavior so this approach would be more useful than host reputation which is entirely reactive and ASN reputation which
does'nt allow for different parts of an organization which might run their networks with different policies.
In our system an ASN's reputation at any given time is based on aggregating up the reputation of all the prefixes originated by that ASN. Therefore it does
not really have a reputation that exists independently of its component prefixes. If an ASN were to be transferred we would simply recompute the
current reputation to be based on the new set of prefixes it is originating.
For prefix reputation you would want to track it on a historical basis with the assumption that it is quite unlikely that a prefixes reputation will undergo a sudden change and therefore tracking historical data would be useful. We do not do this right now, but this is not a solved problem yet ;)
More information about the NANOG