TL;DR: NO, not practical.
Inspired by "Cracking passwords using ONLY response times" video by mCoding.
The scenario:
- There is a web application (written in Node.js, but this applies to any language) which implements some kind of API, which uses a secret key to restrict access to some of its parts.
- The key is stored as plain text, the app compares it with a key provided via an HTTP header (it could be a part of a request body, this is not important). This is a bad security practice in general, but in this case imagine the key is kept secret from end users and potential attackers. For example, it can be only used by another server-side application controlled by the same party (e.g. two microservices talking to each other).
- The keys are compared as strings, using the
===operator, which supposedly works like this (see lib/compare.js for an example "slow" implementation):- First, the lengths of strings are compared
- If lengths match, the first characters are compared,
- If the first characters match, the second are compared,
- ... and so on, until all characters match => the strings are equal
- Otherwise, the strings are not equal.
- There is no request rate limit, or any other mechanism which would prevent an attacker from brute-forcing the app.
- An attacker can perform HTTP requests directly to the target app.
In this scenario, theoretically, a timing attack is possible to crack the key. In practice, however, time delays produces by the string comparison are negligible compared to the delays caused by the network and other sources:
Note that we did not include any string comparison operations in our research. This was intentional, since previous researchers have made it clear that measuring these remotely is almost certainly impractical in the vast majority of cases.
Web Timing Attacks Made Practical (Morgan & Morgan, 2015)
There are studies examining the limits of such attacks:
At what empirical resolution can an attacker time a remote host? The resolution an attacker can time a remote host depends on how many measurements they can collect. Our simulated attacker using statistical hypothesis testing was able to reliably distinguish a processing time differences as low as 200ns and 30µs with 1,000 measurements on the LAN and WAN respectively with.
Opportunities and Limits of Remote Timing Attacks (Crosby et al., 2009)
I wasn't able to achieve the accuracy of 200ns because in this experiment an attacker supposedly has much less insight into the target web application inner structure from what is described in the paper above. However, the limit of 30µs in a real scenario (over the Internet) makes even most advanced timing techniques impractical for performing a timing attack on string comparison. (See benchmarking results at the end.)
- Generate a key:
npm run keygen(Only works in *nix shells: Linux, Mac OS X. For Windows, manually createkey.txtwith a key 1-32 chars long, containing only latin alphanumeric characters: numbers, upper/lowercase letters) - Run a server:
npm run server:unsafe - Crack key length:
npm run crack:length- this will print lines like8 was the slowest X time(s)repeatedly - the number8in this case is the most likely length. The algorithm is probabilistic, it does multiple passes to collect the data to account for systematic changes in round trip times of HTTP requests - Crack the key itself:
npm run crack:key 8(8is the key length from a previous step) - this will eventually guess the key, trying characters one by one.
"Slow" string comparison uses a special sleep method b/w operations - see lib/compare.js, and this slows
down each operation enough so that each additional comparison step adds significant delay, which is noticeable when
comparing round trip times of requests.
The steps are the same, except step #2: npm run server:safe (instead of unsafe)
You will see that step #3 gives random guesses to the key length. Even if you start step #4 with a correct length, it will infinitely produce wrong guesses and start over in a loop.
Run the benchmark:
$ npm run benchmark
> [email protected] benchmark /home/dmitry/projects/node-timing-attack
> node benchmark.js
String comparison: 1453ns
Slow string comparison: 942453ns
HTTP request: 593552ns
Artificially slowed down string comparison works in a similar time compared to HTTP requests in this specific setup, so even the most simple statistical method would be sufficient to perform this kind of attack.
Normal string comparison operation is 2 orders of magnitude faster than a network request, even on localhost, in the app which only checks the key. In a real world scenario, the difference would be 3 or more orders of magnitude. This makes it impossible to perform a timing attack.
This is in agreement with [Crosby, 2009]: string comparison produces delays of only about 1000-2000ns, which is much lower than 30µs threshold described in the paper.
Timing attacks are possible when a target operation is slow enough compared to the noise (network, other operations). [Morgan, 2015] gives a following list of examples (times measured for examples written in Python 3.4):
| Task | Approx. Execution Time (ns) |
|---|---|
| MD5 of an 8 byte string | 6150 |
| Parse trivial JSON string | 13700 |
| Parse moderately complex JSON string | 29300 |
| Parse a 2 parameter HTTP query string | 42900 |
| SQLite memory SELECT | 27800 |
| SQLite on-disk SELECT | 45500 |
| Open and read (but not close) a 1 byte file | 83000 |
Benchmarking methods I've used are based on the process.hrtime() method in Node.js v14. The algorithm of running the
benchmarks does not account for all important factors, but is sufficient for a simple demonstration.
More sophisticated time measurement techniques could exist which would potentially allow me to perform a successful timing attack in the described setup, such as measuring response times by examining network packets instead of benchmarking from the inside of a script. However, I still doubt any of these techniques would be sufficient in a real world setup over the Internet.