Comment by sillysaurus2 | Hacker News Reader

Comment by sillysaurus2 | Hacker News Reader

Parent Full thread

I'm trying to figure out how my understanding of the paper went so horribly awry, but I'm just not seeing it. From the paper:

The first family of attacks, termed synchronous attacks, is applicable in scenarios where the plaintext or ciphertext is known and the attacker can operate synchronously with the encryption on the same processor, by using (or eavesdropping upon) some interface that triggers encryption under an unknown key. For example, a Virtual Private Network may allow an unprivileged user to send data packets through a secure channel. This lets the user trigger encryption of plaintexts that are mostly known (up to some uncertainties in the packet headers), and our attack would thus, under some circumstances, enable any such user to discover the key used by the VPN to protect all users’ packets. As another example, consider the Linux dm-crypt and cryptoloop services. These allow the administrator to create a virtual device which provides encrypted storage into an underlying physical device, and typically a normal filesystem is mounted on top of the virtual device. If a user has write permissions to any file on that filesystem, he can use it to trigger encryptions of known plaintext, and using our attack he is subsequently able to discover the encryption key used for the underlying device. We have experimentally demonstrated the latter attack, and showed it to reliably extract the full AES key using about 65ms of measurements (involving just 800 write operations) followed by 3 seconds of analysis.

Questions:

1) In that example, the attacker "discovered the key used by the VPN to protect all users' packets." Does that mean the VPN reused the same AES key for each encryption?

2) If so, then if the VPN were to change the encryption key each time it encrypted a packet, wouldn't that protect against this attack? The attacker had to be able to trigger 800 encryptions that all used the same encryption key, right? Or does "write operation" refer to something other than "the attacker triggered an encryption of some known plaintext"?

The problem Osvik and Tromer are pointing out isn't that the VPN doesn't rekey; it's that the AES implementation it uses can't prevent a localhost attacker from learning its secrets by watching the footprints it leaves in the cache.

There is no reasonable solution to that problem that doesn't involve eliminating the side channel. The problem isn't the key management; it's the broken AES deployment.

What I think you're saying is, "the paper's point is that side channel attacks can be used to determine the key."

I understand that... that's the whole point of the paper, after all. But wouldn't rekeying defend against this attack? And therefore how is what I originally said wrong (and hence how is it safe practice to reuse AES keys in general)?

(I'm only trying to understand this.)

If you have a side channel attack then a single use of the key would be sufficient to break it-even a OTP would be vulnerable.

Then what does "800 write operations" refer to, if not 800 encryptions? How can a side channel be used to break a single encryption?

Also, how would a one time pad be vulnerable to any side channel attack? That seems like it makes no sense.

What am I missing here?

http://en.wikipedia.org/wiki/Side_channel_attack does a really good job Of describing it, but, in short, a side channel attack isn't a weakness in the algorithm - both a OTP and AES are strong. Rather, the side channel attack describes information leaked in a particular implementation. If your OTP implementation left some registers or cache lines in a state from which key information could be gleaned, then the attacker would be able to decrypt your message. Side channel attacks are about weaknesses in implementations, not algorithms.

Yes, but it makes no sense that a one time pad would leak key information via cache timing, because the entire operation is just 'message xor key'."

With a one time pad, the key's length has to be the length of the plaintext. There's no way for the key to be leaked through cache timing, because no memory location is ever read twice.

The entire operation is simply: read a byte of plaintext; read a byte of key; XOR the two, writing the resulting byte"

It makes no sense to use cache timing attacks against one time pads, because the speed of memory access is independent of the key and plaintext!

If you have a side channel attack then a single use of the key would be sufficient to break it-even a OTP would be vulnerable.

Let me try to put it another way: as far as I know, there are no reasonable side channel attacks against one time pad encryption, and certainly no cache timing attacks. So I don't understand why you're saying OTP is vulnerable to side channel attacks. Also, how is it possible for a single use of the key to be vulnerable to a timing-based side channel attack?

If I'm wrong, please point out precisely what I've said that's wrong, so that I can learn from my mistake.

I'll reword my original questions for clarity: does "800 write operations" mean "the attacker initiated 800 separate encryption operations, each using the same encryption key X, then analyzed the cache timing in order to infer X"? If so, doesn't that mean this attack can be defended against by changing the AES key for each encryption operation? And if that's correct, then by definition isn't it less safe (if not unsafe) to reuse your AES key?

There was a paper just published that built on Osvik & Tromer that recovered a GPG RSA key from a single run of GPG. This problem has nothing to do with key reuse.

I think this is the paper you're referring to: http://eprint.iacr.org/2013/448.pdf

The only reason they were able to recover the key in a single run is because GPG uses branching in its inner loop, making it trivial to determine the difference in timing between 1 and 0 bits of the key. AES requires no such branching. Attacking AES requires that the attacker try to predict the key, then calculate how long encryption would take if the attacker's guess was correct, then compare that guess to how long it actually took. This means it's not possible to recover the AES key in a single run unless the attacker happened to guess it correctly the first time (which is effectively impossible).

In other words, the paper you cited apparently doesn't build off of Osvik & Tromer at all, except in the sense that they're both cache timing attacks. The paper uses a completely different technique than Osvik & Tromer.

So how is it safe to reuse your AES key? I'm just trying to understand. Since the attacker needs multiple runs to attack AES, then wouldn't the logical defense be to rekey after every encryption?

It seems like you've been saying "it's impossible to defend against a side channel attack." But that doesn't seem to be true in the case of AES, as long as you don't reuse your key. What am I missing?

No, that's not what I'm saying. I'm saying that the solution to side channel attacks is to eliminate the side channel, not to redesign the cryptosystem around it.

I keep telling you this: you can reuse AES keys. Reusing an AES key does not in any intrinsic way reduce the security of AES (at least, not unless you encrypt a truly staggering amount of data with it).

However: you should not be working with encryption at this level of detail anyways. Sillysaurus 2, please go download Nacl and play with it. You'll be using 2012 crypto instead of debating implementation flaws in old crypto.

Okay, I was finally able to find a direct answer to my question by studying NaCl.

http://cryptojedi.org/peter/data/tenerife-20130121.pdf

"NaCl systematically avoids all loads from addresses that depend on secret data"

"NaCl systematically avoids all branch conditions that depend on secret data"

So NaCl is an implementation of AES encryption that prevents cache timing attacks. Since there are no other reasonable side channel attacks on AES, the AES key remains secure. Therefore reusing your AES key is ok provided you're using NaCl. (So you should probably be telling people they need to be using NaCl, not merely AES.)

Thank you for the answers. Also, thanks for pointing me at NaCl. That's the most exciting crypto slide deck that's ever been written!

Reply on news.ycombinator.com