On Fri, Jul 14, 2017 at 11:41 AM, Roland Dobbins <rdobb...@arbor.net> wrote: > > On 15 Jul 2017, at 1:01, Melinda Shore wrote: > >> It might make sense to kick it over to ops for a discussion with people >> whose meat and potatoes is monitoring, management, and >> measurement. > > > As someone who is ops-focused, I think this is an excellent suggestion! > > There have been several assertions posted to the list recently regarding > various aspects of security and their intersection with encryption. It may > be useful to take a moment and clarify a few of them. > > With regards to DDoS mitigation as it relates to encrypted attack traffic, > only a subset of attacks in a subset of situations can actually be adequately > mitigated without full visibility into (and often the ability to interact > with) the traffic within the cryptostream. There are various ways to > approach this issue, including full session termination and 'transparent' > detection/classification, the latter of which isn't of course feasible in a > PFS scenario. Each of these general approaches has its advantages and > drawbacks.
DDoS mitigation can be done at endpoints, and indeed has to be given that other tools do not know which endpoints need to be rate-limited or are expensive. A lot of distinct things are being crammed together into DDoS, and the fact is endpoints can deal with application layer attacks via rate-limits, CAPTCHAs, and other techniques, while not-application layer attacks don't require visibility to defeat. What can a middlebox do that an endpoint cannot? Globbing a bunch of distinct things together is not helping the debate. > > Very specifically, fingerprints of encrypted streams are not in fact adequate > for DDoS defense; again, they're only useful for a subset of attack types in > a subset of situations. > > In the case of detecting and classifying hostile activity within a given > network - which isn't limited to malware spreading, but includes data > extraction, attempts at unauthorized access, attempts at subverting > additional devices, et. al. - the same basic caveats apply. It is not in > fact possible to adequately detect and classify all, or even a large subset, > of hostile network traffic without visibility into the cryptostream. There > are some gross behaviors which can be detected/classified whilst standing > outside the tunnel, but assertions to the effect that all or most of what's > required in this arena is possible without visibility (one way or another) > into the relevant encrypted traffic are incorrect. > > It's also important to understand that inserting proxies into multiple points > of a network topology is not cost-free, nor an unalloyed good. It is > impractical in many circumstances, and has highly unwelcome side-effects in > many more, including a negative impact on reliability, performance, and > availability, as well as broadening the potential attack surface. Endpoint > monitoring does not scale well, is often impossible to implement due to both > technical and administrative challenges - and one can't really trust > endpoints to self-report, anyways, as they can be subverted. You are conflating different situations. If you want to detect unauthorized access to a resource, having the resource which determines access anyway log that is enough. Exfiltration detection based on looking for sensitive identifiers doesn't work: real attackers will encrypt the data and dribble it out slowly or pretend to be videoconferencing. Did any exfiltration detector stop heartbleed? Endpoint subversion is not an issue for access control: if the endpoint is subverted, access has been gained. It isn't an issue for DDoS protection. Malware detection on the network does have to worry about subversion of endpoints, and yes, being able to recognize anomalous traffic is helpful here. As for attack surface why is "Press here to get plaintex of everything" not a major, major increase in attackability? Any honest assessment of middlebox techniques would acknowledge they can also be attacked and are inherently unsegmented. Just because many large organizations have decided not to adopt the smart-endpoint, dumb-network model of the Internet doesn't mean the Internet needs to be redesigned to accommodate them. I am seeing middlebox vendors and one vertical express need for this. > > > In many scenarios, one form or another of network-based visibility into > encrypted traffic streams within the span of administrative control of a > single organization is legitimate, vital and necessary. It is not > 'wiretapping', any more than tools such as tcpdump or telemetry formats such > as IPFIX and PSAMP can be categorized as 'wiretapping'. The fact is, the > availability, confidentiality, and integrity of systems, applications, and > networks that everyone on this list relies upon is highly dependent upon the > ability of organizations to have visibility into encrypted traffic streams > within their own networks, for purposes of security as well as testing and > troubleshooting. > > > How this can be accomplished is a matter for further discussion. But it's > important that everyone focused on this topic understands that it is simply > not possible to successfully defend against many forms of DDoS attacks nor to > detect and classify hostile network traffic in the context of encrypted > communications without visibility into the traffic in question, via some > mechanism. The same goes for troubleshooting complex problems. Which DDoS attacks specifically? And if the traffic isn't hitting endpoints, does it matter? > > > Those with operational experience at scale will likely recognize and > acknowledge the difficulties and challenges noted above; others may wish to > consider these factors and their impact on the operational community and the > networks, services, and applications for which they are responsible, and upon > which we all depend, every day. I'm currently working at place you notice when we go down. We do not use these techniques on our production network. We use mutual TLS for a large number of internal services, even within the datacenter and the same machine. We debug many complex issues through logs and specialized tools to examine troublesome state, as well as request tracing. I've not personally had the pleasure of doing this, but I know it is possible because it is done every day. Finally, most software can export the secrets from TLS connections to a file. The capacity being asked for already exists. > > ----------------------------------- > Roland Dobbins <rdobb...@arbor.net> > > > _______________________________________________ > TLS mailing list > TLS@ietf.org > https://www.ietf.org/mailman/listinfo/tls -- "Man is born free, but everywhere he is in chains". --Rousseau. _______________________________________________ TLS mailing list TLS@ietf.org https://www.ietf.org/mailman/listinfo/tls
