Hi, Robert! Hi, Jon!
As usual, please take me not as being "That fellow who is a pompous ass and says things that aren't true" but rather as "that fellow who knows that he is probably wrong about some stuff, and doesn't know a better way to find out what he's wrong about than getting corrected."
IOW, this is what I think now, but if it's not reasonable, I want to thing something else. On Fri, Nov 4, 2011 at 9:01 AM, Robert Ransom rransom.8774@gmail.com wrote:
On 2011-11-03, Jon Callas joncallas@me.com wrote:
Zooko forwarded a hash question over to the SHA-3 competition mailing list, and mentioned the discussion that has been going on here. He's going to forward over comments that I made and John Kelsey made. Nonetheless, I'd like to offer some comments on what I've read in a larger context.
I don't see any discussions of requirements, goals, or anything like that. I see discussions of what hash function to use, ECC, symmetric modes, and so on, and yet no reasons why you would do anything. If you don't have requirements, you aren't going to get a good answer. Any of those proposed solutions needs to have an answer to the question "What is this trying to solve?" at the very least. The next question should be, "How does it solve it?"
For ‘What is this trying to solve?’, see:
On this last point about cell-tagging attacks, see also https://lists.torproject.org/pipermail/tor-dev/2011-November/003024.html for my current thinking.
See also the tor-dev mailing-list threads started with these documents:
Right; and keep in mind that the requirements themselves are also in a evolving draft status.
I think that one of the main problems with the draft I circulated the other week is that it tried to start discussion on three questions:
* Should we keep using our current primitive operations (SHA1, AES-CTR starting at IV 0, 1024-bit DH, 1024-bit RSA, OpenSSL PRNG) or switch to something else? * Regardless of what primitive operations we're using, should we keep using them in the same ways in our protocol, or are their protocol improvements we could reasonably make? (That's the main point of Section 6 for relay crypto, and of the Goldberg-Stebilia-Ustaoglu "ntor" design for circuit exension.) * Assuming that we pick any primitives and protocols other than the ones we currently have, how can we best migrate?
It should probably have been predictable that the first one would have gotten all the attention. Still, I'm not going to call "bike shed" too loudly here, since it helps people get involved in design, *and* the end result is actually important.
Let me try to motivate this and explain goals a bit better. We want to choose our crypto primitives and protocols for a few properties. Some of the more justifiable ones from an engineering perspective are:
[Please treat the above as a draft and tentative set of goals, motivations, and requirements. If you think any of these are stupid and wrongheaded, or just unmotivated, please tell me so!]
A) They should be strong enough that attacking the crypto primitives and protocols is not the best attack on Tor. (We have this property today, I believe, and will continue to have it so long as nobody in the research community comes up with a solution to the low-latency traffic correlation attack.)
B) Further, crypto attacks should not be even _close_ to the best attacks on Tor. (Except for the cell tagging attack, I think we have this property today. With cell tagging, however, the attack is IMO only a little worse than the end-to-end timing-attacks that a smart attacker would be doing instead. Moreover,, to the extent that it's feasible, I'd like to reach a situation where every crypto- or protocol-based attack on Tor is (and remains!) many orders of magnitude more expensive than the traffic correlation attacks that the research community doesn't know how to solve. Ideally, I'd like every crypto- or protocol-based attack to be hideously expensive on an absolute scale, if not downright impossible.)
C) We need good performance. Our original design goal was "commodity hardware should be able to saturate commodity bandwidth." Right now, though, some of the busier servers do seem to be getting CPU-bound. (If I'm remembering what I've seen from profiles, the time sink seems to be to some extent AES and SHA1, and in larger part exponentiation for RSA and DH. But I'd want to double-check that to be sure. See comments below on performance requirements.) We also need our cell-crypto protocol to be space-efficient: bandwidth is at a premium on the Tor network.
D) We'd like a large security margin in our protocol designs. For example, if we have a choice between two protocols, where one of them requires that SHA256 be collision-resistant while the other one only requires preimage-resistance, it would be nicer (ceteris paribus!) to use the second. Or if we can use our cipher in two modes, one of which requires that it resist adaptive chosen plaintext attacks, and the other one only requires that it resist known-plaintext attacks, the second one would seem (again, all other things being equal) better.
E) Because our crypto aims to provide forward secrecy, the parts that do so need to have higher life expectancy than the rest of the rest of the system. For example, it's no big deal if somebody in the future manages to forge signatures from a key that's no longer valid: that doesn't open up new attacks. But if they break a stream cipher or a DH-based key agreement scheme that we are no longer using, that would be poor. That argues (I think) for higher security margins in our key agreement, key derivation, and encryption than in signatures and integrity.
F) We ought to look for DOS opportunities and resist or mitigate them as possible. (This isn't limited to crypto stuff, but it does show up here a bit.)
And here are some implementation concerns: G) As much as possible, we shouldn't have to be the primary maintainers of our crypto primitives.
H) We need portable implementations of our primitives.
I) We need high-quality implementations. To me, right now, this suggests that I'd like categorical rather than empirical arguments for non-viability of timing attacks. (IOW, I would much rather say "This implementation cannot have a timing side channel" than "Any timing side-channel that this implementation might have is probably tiny, infrequent, and hard to exploit.")
And here are some more fluffy concerns:
J) Nothing should be embarrassing, even if it's safe. It provokes unconfidence and uncertainty when there turns out to be a weakness in a primitive or protocol... even if as in D above we give ourselves enough margin to tolerate unforeseen weaknesses in our protocols or primitives, and even if as in B above we ensure that crypto attacks remain a deeply stupid way to attack Tor.
K) It would be nice not to be the biggest or most prominent user of anything if we can possibly help it. This ensures that a wider variety of cryptographers are looking at it,
[...]
Protocols are almost never broken because of the crypto. That's not the problem. Tor as a protocol and as a *service* has many things it needs to worry about long before you should worry about the crypto. Don't lose sight of the real issues. Trust me, they aren't the crypto.
Believe me, we know. Just because this is the Tor development topic that you're seeing first does not mean that it is the only one, or even the highest-priority one on my plate. :)
You need to stick towell-understood ciphers. Do not be frightened by warts on a function. The property of being well-understood means that it will have warts.
I think I either agree or disagree with you here depending on what you mean by 'warts'. I agree with you when it means stuff that's undesirable but easy to work around (like the length-extension properties of Merkle-Damgård hashes), or that is undesirable but impossible to exploit (like a 2^126.5 attack on a 128-bit system).
But I am less sanguine about stuff that weakens a primitive to a degree that I don't understand. For example, consider the AES256 key schedule problems, which as I understand them are interesting but as yet unexploitable, since (again, if I understand right) you need a ridiculous number of plaintexts encrypted with a related keys, and any decent KDF will ensure that you don't have such related keys. But these attacks do seem to get better every year, they make me nervous about AES256. (We're using AES128 these days, as noted elsewhere.)
You're right that it's a major cognitive error to think that a system with fewer known problems is necessarily better than one with more known problems, unless you know how much attention both systems have received.
Additionally, you should stick to well-recognized constructions, too. The main reason is that warts and all, we know most about these things. The alternatives frequently either have different warts, or have unknown warts that will bite you in the ass when they're finally found.
Agreed; we shouldn't be rolling our own MACs, for example.
This is an area BTW where our existing protocols are not so great: there are some places in the protocol where we use H(x||y) when HMAC would be a better choice. Our ersatz KDF is as-yet-unbroken but kinda idiosyncratic. [ H(secret || [0]) || H(secret || [1]) || ... ]. We made some roll-our-own choices when we were getting started, and I'd like to move towards standard and well-analyzed stuff instead.
Also, this is an area where we don't have off-the-shelf solutions for everything. For better or worse, there are not a good set of highly analyzed protocols for some of the things we need to build a working onion routing network these days. Most notably, the circuit establishment protocol and the cell relay protocol are ones where we had to invent stuff as we began, since there didn't seem to be a preexisting solution for either. In both cases, we wound up with something kind of inefficient with some well-known but not-so-bad issues.
And in both cases we wound up with protocols where I believe we could do better today, given care and attention.
Moreover, Tor is an important protocol, service, and system. If you use a weird-ass construction, then a flaw in that will be publicized as a flaw in Tor. If someone finds a non-important problem in AES or SHA256, you'll be shielded by ubiquity. If you use the Whizzy2^n-1 whatever, then when a problem is found in it, that will hit the press as "FLAW FOUND IN TOR! EVERYONE RUN FOR THEIR LIVES!" and you'll be left defending it yourself. The crypto elders will also cluck their tongues and say, "Yeah, they really should have used what NIST documented in SP 800-9999." Lastly on that, Tor is no important that you'll be a target. People *will* target Tor because a teenie, tiny, unimportant break in Tor is a g uaranteed paper acceptance in Usenix Security or Eurocrypt. Anonymity loves crowds; so do cryptographers.
And how!
Tor's use of well-recognized constructions hasn't stopped the recent wave of ‘semi-technical bullshit attacks’ and plagiarisms of old, not-particularly-effective attacks from winning the attention economy.
(One of the claims in the latest anti-Tor publicity campaign is that using AES in CTR mode is a ‘bugdoor’, and that we knew that we should have used CBC mode instead. This is even after the news that use of block ciphers in CBC mode breaks SSL's security. Yes, people *lie* in order to get publicity for an alleged attack on Tor.)
I don't think that's a lie -- I think that's a combination of a probably-independent rediscovery of the cell-tagging attack with an inadequate understanding of Tor's threat model and the problems with CBC for our use.[*] Never attribute to malice etc.
(And never attribute to plagiarism what can be explained by an unfamiliarity with parts of the literature, or a buried-but-still-present credit. The more serious the charge, the more caution and forbearance is warranted, especially when people are already ticked off with one another.)
[*] The problems with "let's use CBC" in summary: If you do CBC wrong and don't do a new IV per cell, you're exposing yourself to something like the Rizzo-Duong attack on SSL. If you do CBC right, you need one IV per layer per cell, which is a pretty high overhead for Tor. CBC doesn't get particularly good tagging resistance: it just lowers the number of bits you can tag and your ability to recover the stream afterwards. So if you're willing to burn one block per hop per cell, you're better off using those bytes for a real MAC or GCM or something than for a CBC IV. [And if you are hoping to avoid transmitting IVs by having each cell get a fresh set of IV derived from a keyed PRNG or something, you aren't really doing orthodox CBC any more.]
Here are some specifics:
- Hash functions.
[...]
Skipping this part for now. FWIW, I like Skein myself, and actually like its "pile of features" approach. (To my mind, if you don't provide a construction in a hash function for doing X (where X is personalization, MAC, PRNG, or whatever), people will go and do it badly. I know I have!) But you're right that we ought to write up speed requirements first.
I think that for all our speed discussions, we ought to stop, instrument some busy servers, and look at how many cell encryptions, circuit extension handshakes, connection handshakes, and miscellaneous "other" crypto operations they actually spend their time on. Then we can talk categorically about which parts of the system matter how much for performance.
We do need to be reasonably good on weak clients, but it's more important to be fast on servers IMO.
[...]
- ECC. I don't see a reason why you're looking at it; my remark isn't code
for thinking it's a bad idea, it means I don't see *reasons*, and yeah, yeah, I keep harping on that.
For me, the big reason is this one:
- We want forward secrecy for circuits, so we use Diffie-Hellman, too.
For every circuit extended to a relay, the relay must perform two scalar-multiply operations. Currently, we use 320-bit exponents in a 1024-bit mod-p group published in the 1990s.
Because of reasons C (gotta be fast), D (gotta have security margins), and E (forward secrecy needs the biggest margins) from above, I think our DH operations are the ones where we should be most conservative about key size, and that plus performance requirements suggests a move to ECDH to me. Whether it's next year, in two years, or in three years is going to depend mostly on what the lawyers say.
[...]
Should you go to ECC, use P-256; do not use some variant of 25519. You want to be in a crowd. You want to use something that lots of smart people have looked at. That means P-256. For *any* counter argument, please supply the requirement or goal that makes P-256 the inferior choice, preferably with a metric. Absent such, P-256 is your baby.
Possible! I want to talk to a bunch of ECC people here.
Robert: from your list of properties, I think some of them are desirable, but not flatly compelling on their own. Nothing in ntor will _fail_ if we need to do group membership tests. Smaller keys and point representations are nice-to-have, but not so critical on their own.
The overall speed advantage is what has been most attractive to me, but before we can decide whether it's compelling, we need more analysis and data.
I think we can reach the best conclusions here by avoiding getting too married to any particular solution.
[...]
- Stream ciphers. For a stream cipher, you should be doing at AES-CTR.
Really. Stop worrying about the warts. There are hardware implementations with AES-NI, or attached hardware that don't have problems. ArmV8 has AES in hardware along with SHA256, which another reason to stick to the well-understood. AES-NI on an x64 processor runs at under 1 clock per byte. That's all goodness.
I don't have one of those. Why should I have to use a cipher which no one has implemented securely *and* efficiently on my computer? Why do you want to continue to inflict side-channel leaks on our users in oppressed countries who have even older computers?
I doubt Jon wants to inflict anything on anybody.
At their best, hardware-based AES implementations can be pretty good these days. We do however need to see whether we can find better software-based implementations than openssl's, which is not making me thrilled at the moment. (Its solutions for avoiding cached-based timing seem kludgier than I would like, if I am understanding them correctly.)
For software AES, some of what I've reading about bit-slicing implementations AES is pretty impressive. Assuming that we use as much of any given counter stream as I believe we typically do, we can easily tolerate the requirement of having to encrypt 1-4k of stream material in a go.
Of course, there could be issues here that I don't at all know about! More knowledge and analysis is needed.
[...]
In general, if we don't have to write a paragraph or two explaining why the published attacks on an algorithm used in our relay crypto are bogus or irrelevant to Tor, we're using too many rounds of it. The relay crypto only needs to make cryptographic attacks harder enough than the end-to-end timing correlation attack that no one will ever benefit by implementing an attack on our crypto. Anything we are willing to use will far exceed that goal.
I think some of the goals I list above suggest that we want more of a security margin that you suggest here. In particular, consider that the security of our cipher affects our forward secrecy. (Not that you're necessarily wrong, but it is not 100% clear-cut to me.)
[...]
- Other comments. In the relay crypto comments, there was something about
counter "in a less scary mode" -- just use what they describe in NIST SP800-whatever.
Guessing that you mean SP800-38A appendix B? There's more than one thing that looks like counter mode or a variation thereof in the SP800-... series.
Assuming we're still on counter mode, yeah, what you said. Right now we use counter mode with a fixed initial IV, which isn't *bad* necessarily, but random-IV would be better.
Ugh, what a long mail! I hope we are closer to asking the right questions soon.
cheers,