Draft: draft-mm-wg-effect-encrypt-09 Date: 2017-04-07 Overall This document is unfocused and unclear in its intent. It is filled with unsubstantiated claims, misleading statements, bias, and implied recommendations for bad practices. I would oppose its publication in the current form. Furthermore, I don't believe that small scale or editorial changes would be sufficient to correct these problems. Aside from weakening its utility and/or arguments, the lack of focus will allow this document to be abused in various ways. For instance, it might imply IETF endorsement of the practices described in the document. Given that the IETF has published positions that in some cases reject these practices, this document needs to be very precise with its claims. In its current form, it is not nearly careful enough. If this document is going to be published as an RFC, then it needs to be very clear about its intent and its position - or lack thereof - toward these practices. It might be possible to say that these practices were employed in networks in the past and that encryption is reducing the viability of those practices. A value-neutral statement like that might be acceptable as long as it were framed carefully. What is problematic is the implicit argument that is presented. This document could easily be construed as the IETF legitimizing these practices. This includes things on which the IETF has published statements (see for example RFC 2804). This steps firmly into to sticky mess that is the politics of encryption policy. If the IETF is going to make a political statement with this document, then that statement needs to be a lot better than this. Furthermore, any statement needs to be consistent with previous statements, where in this case that primarily means RFC 7258. Right now it reads as an attempt to dilute RFC 7258. If this document were more clearly formulated as an even-handed treatment of what happens today - without judgment - then it could be useful. Here the model of RFC 7754 is a good one. It describes the different facets of a use case, then the different practices that might be employed to achieve that goal. For each practice the pros, cons, and technical limitations are covered with an eye to all involved parties. The Real Purpose of this Document I think that this is the real question that this document wants to ask is: In a world where we are forced to defend against pervasive monitoring by motivated and malicious actors, can we find a role for less pervasive monitoring by well-intentioned entities? There is an interesting discussion to be had here, but I'm convinced after my review that this document is the wrong vehicle for that discussion. I'm also inclined to suggest that this is the wrong question to ask in the first place. To the extent that we have the tools necessary to protect against pervasive monitoring, we have to accept that more-legitimate uses of monitoring are collateral. Mostly, that's just a function of the limitations of the imperfect tools we have. The value in this document is in the rather comprehensive collection of use cases. Sure, some we might not attribute much weight to, like wiretapping. Others are frankly frightening and not just in the Orwellian sense. The problems that are implied by the current practices of network operators is a large vein of unexplored work for the community. Spending effort on finding better solutions to legitimate problems is worthwhile. In other words, the right question might be: For each of these practices that we see today, are the use cases that motivate them legitimate and how might be implement alternative techniques that properly respect privacy and security? Two- and Three- Party Interactions A great many sections of the document are unclear about the relationships between different actors. It seems like many of the activities describe, which are presented here as being beneficial, are performed by a third party without the consent or involvement of the two communicating peers. After all, those are the cases where encryption most often interferes. The assertion that some benefit is accrued to communicating peers as a result of these actions is the constant subject of the text. I observe that in most of the cases listed in this document, the benefit is accrued primarily to the third party. Clarity around the arguments that the document makes with respect to benefits to each of the involved parties is sorely lacking. In addition to this, where benefits truly do accrue to communicating peers, technical options that don't rely on privacy-invasive techniques are routinely dismissed in an offhand fashion. In this way, any moral imprimatur attributed to the use case is used to justify the privacy-invasive method. Document Structure This document is very hard to read (and review). Probably my biggest complaint is that the sections are haphazardly organized. They contain a mixture of use cases (what does SP X want to do) and techniques (how do they do these things). Material is duplicated between sections. Sections cover 10 different use cases while talking about techniques, then other sections talk about different techniques whiles talking about a scenario. This leads to a lot of repetition. There are three axes that this document contemplates: - the techniques that are currently used in networks - the use cases that motivate these techniques - the deployment scenarios in which these use cases manifest It would appear that the document attemps to start from the deployment scenarios. However, this causes use cases and techniques to be mixed. Judicious use of cross-referencing might have made this arrangement tractable, but there are virtually no cross references. Section 1 I find the introduction of this document difficult to parse. I think that what it wants to say is pretty straightforward, but it manages this in quite an obtuse way. Here's what I think that it is trying to say: 1. The IETF (and the larger community) has reacted to revelations of pervasive monitoring by increasing the use of encryption. 2. That has been somewhat successful. 3. More encryption has an impact on some practices that network operators have become accustomed to employing. Expressed in this way, you can see how you could make a value-neutral statement. On the first, you can definitely make an argument based on the documents that the IETF have published. However, it is a mistake here in assuming that the IETF - or the revelations of global-scale surveillance - is directly responsible for the uptick in adoption of cryptographic protection for Internet traffic. This is something that the community as a whole has been working towards for many years; these trends predate the actions this draft focuses on. This is particularly true for the hosted provider cases in Section 3, where business motivations for protection are far stronger than any concerns over government surveillance. Yeah, that's a subjective view, but I'm just reviewing, I don't have to write a statement that will be labelled as having IETF consensus. The immediate focus on opportunistic security is highly misleading, especially in the web case. The amount of opportunistically secured web traffic is miniscule by global standards. Since Firefox is the only browser to support OS for the web, so adjust the following by market share: <https://mzl.la/2oc46ch>. I understand that mail is more often opportunistically secured, and you will hear the virtues of OS sung from the rooftops, but all evidence suggests that this is just a transitory state. Instant messaging is closer to the web case, with the trend toward strong protections that include authentication. Of course, I don't believe that mail or instant messaging represent any significant traffic volume, so you need to be very careful about the nature of the claims that are being made. It has to be that the volume of mail has to be a rounding error when it comes to capacity planning for networks. On the second point, when making claims about prevalence of encryption, I would advise caution when citing statistics. Here's the graph for the Mozilla figures that I think are being cited <https://mzl.la/2obZjrb>. Here's a different graph with a different number <https://mzl.la/2oc263J>. This highlights the point that statistics need to be carefully cited, because bare claims are difficult to assess. Methodology matters when it comes to these sorts of claims - are we talking proportions of packets, requests, flows, or something else? Furthermore, the citations in the document are made more difficult to assess by being measured at very different times (where the two I cite here were made at roughly the same point in time). The statistics do support the notion that there is a significant amount of encryption in use, but it would appear that the claim is that the amount of encryption is *increasing*. This is not an extraordinary claim, but no evidence is offered in support of the claim. It isn't a certain thing either. From the statistics that I have access to, there was a definite uptick in October of 2015, but the rate appears to be stable since then. It appears that the remainder of the document is intended to address the third of my points in some detail. That's a good structure in theory, but see above. I don't think that the "service provider" taxonomy works for this document. Based solely on the structure of the document, the only service provider that this document concerns itself with is the one who forwards the packets. That application service providers are involved as one of the endpoints is relevant in some cases, but not all. For that reason, I prefer "network operator" and to use specific terms like "mail host" or "web server" as appropriate to the context. Section 2 (top section) I like that the attacks on SMTP by network operators is highlighted here. There's an implicit argument here that runs counter to established IETF consensus. Some methods used by service providers are impacted by the use of encryption where middle boxes were in use to perform functions that range from load balancing techniques to monitoring for attacks or enabling "lawful intercept", such that described in [ETSI101331] and [CALEA] in the US. This fails to remain neutral. This implicitly makes a value-judgment about relative morality/acceptability/value of certain practices. What is particularly insidious about this particular form is that it invites interpretation that is subjectively coloured. For instance, as a non-EU and non-US citizen, I might consider the use of the particular lawful intercept techniques offensive; thus I might interpret this as a spectrum between inoffensive to offensive. An employee of the US government might view this somewhat differently and interpret this as a scale between low-value to high-value. I'm guessing here, but this is really just to highlight my point about care. A more serious concern is this statement: The loss of access to these fields may prompt undesirable security practices in order to gain access to the fields in unencrypted data flows. I've read this argument before (see https://queue.acm.org/detail.cfm?id=2508864 for the long form), and it might even have an element of truth to it. However, it's not a statement that I personally attribute any real credibility to, and nor do I think that it needs to be credited with any amount of seriousness. I certainly disagree with the IETF making any such statement. It says that we are collectively willing to bow in response to (anticipated) threats of violence. Section 2.1 This is one of the few sections that talk about what it means to operate a service as opposed to operate a network. Overall, this section doesn't work particularly well. The distinction between integrated and standalone load balancing is an interesting division, but it doesn't leverage this distinction well. What causes a standalone load-balancer to be necessary? Is this something a network operator uses? I see text on NFV later, but it's far removed from the original text and it seems aspirational rather than concrete. On the other hand, many of the concerns in this document simply don't apply to an integrated load balancer. The amount of text on QUIC here is surprising. Given how much of QUIC is changing right now, we shouldn't publish a document that attempts to make claims about what QUIC is or how it is operated. This attempts to dive into the details of QUIC connection migration in a way that presumes much about the outcome of issues that the QUIC working group is still struggling with. I would strongly recommend removing QUIC-specific language from this section and the document as a whole. We have an operational document in the QUIC working group that would be a good venue to discuss some of these concerns. Is this an oblique reference to (the much-loved) RFC 7974? Current protocols, such as TCP, allow the development of stateless integrated load balancers by availing such load balancers of additional plain text information in client-to-server packets. Otherwise, I don't know what it means. BTW, I like this parenthetical: (That said, care must be exercised to make sure that the information encoded by the endpoints is not sufficient to identify unique flows and facilitate Persistent Surveillance attack vector.) I'm going to take this to the QUIC working group, because QUIC certainly doesn't meet that bar right now. It fully facilitates Persistent Surveillance in its current form (proper noun? I haven't really seen that term before, even though it draws a neat parallel with Pervasive Surveillance). Editorial: * "pop" is a term of art that needs explanation. * "QUIC?s", "network?s" - avoid smart quotes, and - more generally - the possessive form for the inanimate. * "QUIC's server-generated flow IDs" -> "QUIC's connection IDs". Section 2.2 I think that this sums the situation up reasonably well. It fails the value-neutral test in a few ways though. As I understand the situation, surveying operates at many levels. For accurate planning, models of endpoint behaviour are used to determine things like whether loss or congestion is affecting throughput. Really good models require knowledge of what users are attempting to do and the applications they are using (as mentioned, things like browser version matter in these models). For instance, it isn't enough to understand that the user is trying to receive 720p video, you also have to know that a 4k stream for the same content is available. The degree to which these models are privacy-invasive is not even acknowledged. Nit: Did you realize that "well-intentioned" is a euphemism? I'd expect that subtlety to be lost on many readers though. Section 2.3.1 I find this sad, even if I can recognize the truth of it: A monitoring system could easily identify a specific browser, and by correlating other information, identify a specific user. As a browser vendor, we don't consider this to be a feature, it's a bug. Section 2.3.2.1 This is a strange section under which to put caching. Caching is a use case akin to compression. I don't see it as belonging to the class of things that rely on DPI. You just intermediate the cleartext HTTP in the way that RFC 7230 recognizes. Note that RFC 7230 explicitly does not endorse the practice of transparent proxying for a range of reasons, not the least of which is the complete lack of transparency (yep). Section 2.3.2.2 Again, it's strange to see differential treatment under DPI. I was fairly sure that fingerprinting was used here as well. Section 2.4 This section doesn't even attempt to recognize that applications are much better now at scaling content to suit the device on which that content is intended to be consumed. It should. This section should not represent compression as an unmitigated good. I'm told that significant damage can be done to video streams by "well-intentioned" compression middleboxes. Yet again I see a call back to the implicit threat of violence in "they will adopt undesirable security practices". Statements in that form have no place in IETF RFCs. Section 2.5 This mentions RFC 7754 then blithely ignores its primary conclusion. Generally, The treatment in RFC 7754 of this subject is more even-handed. The use of particular examples (betting, gambling, dating) shows cultural bias. Jargon warning: "core network", "the mobile network" ("the"?) Section 2.5.2 This type of granular filtering could occur at the endpoint, however the ability to efficiently provide this as a service without new efficient management solutions for end point solutions impacts providers. The initial premise here (up to the first comma) is the conclusion of RFC 7754. I disagree with the conclusion, and I suspect so does RFC 7754. In scenarios where this matters, endpoints are routinely managed centrally. The range of options for this sort of management are plentiful and diverse. Section 2.5.3 This describes a captive portal, which in the case where the service provider has a relationship with their customers, seems like a poor solution. See the CAPPORT working group for ways in which people are actually working on a solution to this problem. Section 2.6 The capitalization of section headings is inconsistent here (and throughout). Section 2.6.1 Isn't this a duplicate of Section 2.1? Section 2.6.2 I can't parse this statement: Approved access to a network is a prerequisite to requests for Internet traffic - hence network access, including any authentication and authorization, is not impacted by encryption. And then we get into zero rating, which is a hot-button topic. No acknowledgment given to the sensitivities on the subject, or even a superficial exploration of the technical options that are available. This text references a non-existent Appendix. Section 2.6.5 This paragraph is strange. It starts by describing what appears to be use cases (are those scare quotes?). It finishes by implying that these are attacks on privacy (which they are). What is "Non-Customer Proprietary Network Information"? I hope that this isn't an attempt to somehow privilege the information so that it seems less than privacy-invasive. Section 2.7 This is one of the areas that is most legitimately affected by increased adoption of encryption. I have no doubt that having detailed information about the application and its use makes the process of troubleshooting easier. This is the first mention in the document of network-based optimization. Was the section describing that cut from an earlier draft? The use of websockets as a primary example is an odd one. It's just not that common. For reference, we see >20% failure rates on encrypted websockets and >50% failures in the clear; no one in their right mind would rely on websockets. If the point is that HTTP is carrying more traffic (see http://conferences.sigcomm.org/hotnets/2010/papers/a6-popa.pdf) saying just that would be more effective. Section 3 It would help if "hosted environments" was defined before diving in. Section 3.1 The paragraph on DLP is ironically amusing. Encryption exists to prevent private information leakage, but the assertion here is that encryption is hampering DLP attempts. It's true that protection of intellectual property is an important business goal, but this handling suggests an architecture that I'm fairly sure is counter to established wisdom (see "SSL added and removed here :-)": https://www.washingtonpost.com/rf/image_404h/2010-2019/WashingtonPost/2013/10/30/Local/Images/GOOGLE-CLOUD-EXPLOITATION1383148810.jpg). Worse, this casually dismisses a model that is actually more likely to work. Malware encrypts too (https://arxiv.org/abs/1607.01639). Section 3.1.1 Who is the customer here? Why does the hosting provider need to monitor access? They have the ability to limit accesses, but this is suggesting that what happens within the envelope of what they permit is important for them to know. I'd be surprised if you could show that this is true. Hosting providers - in my experience - value their ongoing business and would not jeapardize it by snooping on their customers. It's different if the customer opts in to a security service, but that demonstrates a cooperative situation, where the rest of this document is largely concerned with adversarial uses of encryption. Section 3.1.2 This is another example where the organization of the document could be improved. Again with the "SSL added and removed here :-)" recommendation. PCI don't permit that for good reason. Section 3.2.1 I don't believe that monitoring of these types of application require cleartext access. There are two parties involved: the provider of the application and the user of it. The provider has most of the power here and can theoretically design the system to allow whatever level of access they require. To the extent that the user needs protection from the application provider, that is a matter of negotiation between them. I don't see how the adversarial three-party situation presented in the rest of the document applies in this situation. Section 3.2.2 No real mention of the effect of encryption here, or is this just saying that there is no issue here? I don't see the point of the PGP/Dark Mail paragraph. I'd be leery of the IETF saying things like "PGP may be a front runner" though. Section 3.3 There's certainly a lot of detail here, but I had a really hard time extracting value from this section. I can't see how data storage is fundamentally different to the general case of an application provider. Section 4.1.1 I appreciate the recognition that endpoint-based techniques work. The implication that this isn't the obvious solution, much less so. Section 4.1.2 I'm confused about the purpose of this section. This seems to be talking about the sort of network-based methodologies an application provider might employ to ensure that their application is working as intended. Why would the application provider not have access to detailed logging and usage metrics? As an aside, I skimmed through ietf-ippm-6man-pdm-option, since it is cited no fewer than three times in the same paragraph. It makes some fairly bold claims about its effect on privacy that I don't believe will hold up to analysis. Section 4.1.3.1 This section correctly identifies the issue as one of shortcomings in logging, not increased use of encryption. I would have stopped there, but the section persists and ultimately references RFC 7974, which is widely recognized as a monstrously bad idea (the IESG note is fairly clear on this point). Section 4.1.3.2 "HTTP/2", not "HTTP2". I don't see how this section belongs in this document. The 1:1 correlation between actions and flows in HTTP was a mistake of the 1980s that we've spent a lot of time on correcting (sometimes unsuccessfully, see pipelining). Section 4.1.3.3 It's not a "service call" it's just a request. Section 4.2 If the inclusion of a reference to RFC 7457 is to support a claim that TLS is attacked, that's sailing far too close to an endorsement of attacks than I am comfortable with. However, I don't believe that any of the listed attacks remain viable in modern applications. What is most often the case is that a trust anchor is installed on enterprise users' machines and new certificates are minted as needed. In other words, the host that is accessing the HTTPS site is owned by the enterprise and modified so that it can comply with its policies. When you recommend attack like this (to be clear, I would oppose publication of text that does this), you need to acknowledge the downsides. For example, MitM devices routinely break connections, they hide the true security status of communications from endpoints and users, they frequently implement weaker versions of protocols, they often don't include the same degree of rigor in things like certificate validation, and probably many more things that I can't casually list off the cuff. The discussion of caching here warrants a new section. Section 5.1 s/effect/affect No complaint here, just an advertisement for technical solutions that aren't affected by increases in encryption. Good, though I'm not sure if the section meets the document's criteria for inclusion. Section 5.3 No cititation for APWG. Not an IETF working group from what I can see. There's a presumption here that administrators need to perform these tasks. Why can't endpoints do this? It would certainly be a whole lot less invasive that way. Take a look at how browsers implement checks for "bad" sites (which include phishing sites). These methods have some fairly significant privacy safeguards without compromising on performance. Section 5.5 See my comment about about endpoint-based methods. Section 5.6 I believe that the spoofed source address problem is not relevant to this document. It's a fairly well understood problem. That said, if we could wave a magic wand and get BCP 38 deployed, that might be nice. Section 6.1 As this says, it's fairly well understood that - for HTTP - SNI is used. The point that it is optional is interesting, but doesn't deserve the amount of attention this section devotes to it. The clients that don't include SNI are in a diminishing minority. That said, this doesn't pay any attention to another feature in HTTP/2: connection coalescing. Section 6.2 ALPN will be encrypted in TLS 1.3. Section 6.3 This reads as a invitation to perform traffic analysis (a statement that I oppose). Note that we have added padding to most recent protocols (HTTP/2 and TLS 1.3 in particular) to give endpoints the ability to resist this sort of attack. Note that block ciphers can add ~240 octets of discretionary padding per record. That can be pretty effective if you are careful. Section 7 This section is duplicative of much of the rest of the document. I realize that editing is hard, but see my earlier comments about structure. It's obvious that this section is about QUIC. That shows in several ways. It's nice how it doesn't say so out loud, but it leaks through in several confusing ways. See below. Section 7.1 This section is purportedly about encrypted ACKs, which won't happen until we get to QUIC. But the points regarding proxies (b, c, d) are not prevented by encrypted ACKs, they are prevented by encryption more generally. Point e is defeated by integrity protection of ACKs, not confidentiality protection. I hope that the IETF never publishes draft-dolson-plus-middlebox-benefits; it makes claims about the benefits of specific solutions for different use cases with the goal of justifying those solutions. At the same time, it fails to recognize the existence of alternative, often superior, solutions for those use cases. In other words, it has many of the same issues as this document. FWIW, also be OK if draft-thomson-http-bc never went anywhere as well; I don't know how to close the gap between the privacy assurances I wish it had and the privacy weaknesses it has. The phrase "trusted proxies" is a dangerously misleading phrase. The concept of trust is - particularly in this context - frequently abused. Sections 7.2 and 7.3 I'd be interested in learning about the justification for these features (again, go back to my comment about to whom the benefit accrues, and I'd advise caution not to make the trickle-down economics mistake of using second-order effects). Personally, I'm not that sad that they are going to be negatively affected. Section 7.2, Point d seems to be saying something about a multiplexed protocol, not one that has an encrypted transport header (see above regarding QUIC). All these points equally apply to HTTPS, which suggest that encrypted transport headers is not the issue (unless by transport we mean HTTP). Section 7.4 s/GPSR/GPRS/ ? Section 8 I generally agree with the first statement here, but after having read the document, I could take away some very different views on what the "problems at hand" actually are. I suspect that different people will have very different takeaways. The conclusion section of a document is not the right place to start introducing new information, particularly information relevant to RFC 2804. Not sure where this is going: Terrorists and criminals have been using encryption for many years. Because it's disconnected from the remainder of the paragraph. If you are going to invoke the T-word, you had better have a strong argument to make. On the final sentence of that paragraph, the subject ("This") is unclear.
