4 ms·

In general, can timing attacks on a vulnerable implementation be prevented if sonething else guarantees the operation takes constant time? For example, if I ad

by 201984 2y ago

In general, can timing attacks on a vulnerable implementation be prevented if sonething else guarantees the operation takes constant time? For example, if I add a sleep after a non-constant-time operation to ensure it always takes 1s, would that prevent timing attacks? At least with this you can use a faster algorithm, and let the OS do useful work with the rest of the time.

some_furry 2y ago
https://security.stackexchange.com/a/96493 https://security.stackexchange.com/a/96493
- ghostpepper 2y ago
  Unless I misunderstood (which is entirely possible, not a crypto expert) that article is answering a different question: whether adding random delays can thwart timing attacks. The GP was asking about implementing a constant-time delay.
  - some_furry 2y ago
    Ok, let's assume you know how long an operation is supposed to take, but want to stop it from leaking if it returns faster. To accomplish this, you would need a high-resolution timer that runs before the operation, after the operation, and then you can subtract the two and wait. And maybe that might work. But you're not running in kernel space. The resolution of timing information you'd get from this tactic would be noisy, and may reintroduce the signal you were trying to hide. Especially if you're running on multiple threads or physical processors, and aren't careful enough to constrain the operations and timings to the same core. All it will do is slow your software down and maybe not actually help. And that's the best case scenario! Better to just use constant-time algorithms.
    - 201984 2y ago
      Yeah, that's pretty much what I was asking. start = gettime() variable_time_crypto_operation() end = gettime() sleep(1sec - (end - start)) // Use a loop if necessary here If the operation takes a couple ms and the sleep was 1sec, then how much information would realistically leak here? Sure the sleep might not be perfect, but I'd imagine the actual timing difference would get lost just in how much longer the sleep is.
      - theon144 2y ago
        Somewhat tangential, but there are much better options if you're looking for opportunities for optimization. It's literally trying to improve efficiency by skimping on safety features, like trying to save on vehicle weight by removing unnecessary seatbelts or crumple zones. Eliminating side channels concincingly is very difficult, you're just better off taking the tiny performance hit and virtually* eliminating that vector instead of trying to come up with a novel low-density seatbelt. (I say virtually, because even constant time crypto isn't bulletproof - GoFetch, a recent Apple M-series CPU vulnerability inadvertently broke the "constant" part because of a quirk of the prefetcher. Side channels are hard, no need to make it harder.)
    - ghostpepper 2y ago
      is a constant-time algorithm just an algorithm that completes the entire calculation before returning even if it knows the correct answer sooner?
      - some_furry 2y ago
        "constant-time" isn't a literally technically correct descriptor, it's a short-hand What matters is that the variance of execution time, which memory addresses are accessed, etc. is independent of any secret inputs.
        ghostpepper 2y ago
        so theoretically using timers and adding sleeps could solve the problem but practically there are better ways to achieve the goal of no timing information leakage?
        some_furry 2y ago
        Yes. I wrote this a few years ago as an introduction to how to write constant-time algorithms https://soatok.blog/2020/08/27/soatoks-guide-to-side-channel-attacks/ https://soatok.blog/2020/08/27/soatoks-guide-to-side-channel...
thadt 2y ago
Theoretically? Sure. By holding your response to a fixed deadline, you're still using a variable time operation under the hood, but turning it into a fixed time operation, being careful not to leak that timing information to an external observer. So, no problem. Practically? It's a whole lot harder. Having done exactly this in a non-cryptography related situation (sleeping a timer while trying to get time bins as close as possible to a fixed size for binning events without special hardware), you run into all manner of interesting sources of variability. Using hardware is often far simpler if it's available. Back in a cryptography context, it is exponentially harder yet because you now have to worry about what other side channels are exposing your timing information. For example, if you're "letting the OS do useful work with the rest of the time" - what is it doing? If it's off doing anything else that I can observe, then I can time that instead. And we're back to square one. Eliminating side channel leaks is entirely non-trivial.
- 201984 2y ago
  > If it's off doing anything else that I can observe, then I can time that instead. And we're back to square one. Wouldn't these kinds of leaks always be present though? The scheduler can swap your process out at any time, regardless of whether it explicitly calls sleep() or not, so even with a constant time algorithm you'll have some variability.
  - some_furry 2y ago
    The problem isn't variability in total. The problem is variability based on a secret input.

by 201984 2y ago

some_furry 2y ago

https://security.stackexchange.com/a/96493 https://security.stackexchange.com/a/96493

ghostpepper 2y ago
Unless I misunderstood (which is entirely possible, not a crypto expert) that article is answering a different question: whether adding random delays can thwart timing attacks. The GP was asking about implementing a constant-time delay.
- some_furry 2y ago
  Ok, let's assume you know how long an operation is supposed to take, but want to stop it from leaking if it returns faster. To accomplish this, you would need a high-resolution timer that runs before the operation, after the operation, and then you can subtract the two and wait. And maybe that might work. But you're not running in kernel space. The resolution of timing information you'd get from this tactic would be noisy, and may reintroduce the signal you were trying to hide. Especially if you're running on multiple threads or physical processors, and aren't careful enough to constrain the operations and timings to the same core. All it will do is slow your software down and maybe not actually help. And that's the best case scenario! Better to just use constant-time algorithms.
  - 201984 2y ago
    Yeah, that's pretty much what I was asking. start = gettime() variable_time_crypto_operation() end = gettime() sleep(1sec - (end - start)) // Use a loop if necessary here If the operation takes a couple ms and the sleep was 1sec, then how much information would realistically leak here? Sure the sleep might not be perfect, but I'd imagine the actual timing difference would get lost just in how much longer the sleep is.
    - theon144 2y ago
      Somewhat tangential, but there are much better options if you're looking for opportunities for optimization. It's literally trying to improve efficiency by skimping on safety features, like trying to save on vehicle weight by removing unnecessary seatbelts or crumple zones. Eliminating side channels concincingly is very difficult, you're just better off taking the tiny performance hit and virtually* eliminating that vector instead of trying to come up with a novel low-density seatbelt. (I say virtually, because even constant time crypto isn't bulletproof - GoFetch, a recent Apple M-series CPU vulnerability inadvertently broke the "constant" part because of a quirk of the prefetcher. Side channels are hard, no need to make it harder.)
  - ghostpepper 2y ago
    is a constant-time algorithm just an algorithm that completes the entire calculation before returning even if it knows the correct answer sooner?
    - some_furry 2y ago
      "constant-time" isn't a literally technically correct descriptor, it's a short-hand What matters is that the variance of execution time, which memory addresses are accessed, etc. is independent of any secret inputs.
      - ghostpepper 2y ago
        so theoretically using timers and adding sleeps could solve the problem but practically there are better ways to achieve the goal of no timing information leakage?
        some_furry 2y ago
        Yes. I wrote this a few years ago as an introduction to how to write constant-time algorithms https://soatok.blog/2020/08/27/soatoks-guide-to-side-channel-attacks/ https://soatok.blog/2020/08/27/soatoks-guide-to-side-channel...

thadt 2y ago

Theoretically? Sure. By holding your response to a fixed deadline, you're still using a variable time operation under the hood, but turning it into a fixed time operation, being careful not to leak that timing information to an external observer. So, no problem. Practically? It's a whole lot harder. Having done exactly this in a non-cryptography related situation (sleeping a timer while trying to get time bins as close as possible to a fixed size for binning events without special hardware), you run into all manner of interesting sources of variability. Using hardware is often far simpler if it's available. Back in a cryptography context, it is exponentially harder yet because you now have to worry about what other side channels are exposing your timing information. For example, if you're "letting the OS do useful work with the rest of the time" - what is it doing? If it's off doing anything else that I can observe, then I can time that instead. And we're back to square one. Eliminating side channel leaks is entirely non-trivial.

201984 2y ago
> If it's off doing anything else that I can observe, then I can time that instead. And we're back to square one. Wouldn't these kinds of leaks always be present though? The scheduler can swap your process out at any time, regardless of whether it explicitly calls sleep() or not, so even with a constant time algorithm you'll have some variability.
- some_furry 2y ago
  The problem isn't variability in total. The problem is variability based on a secret input.