How can I “throttle logins” or “how can I rate-limit login attempts” is a common question raised from time to time among web application developers.
In this post I describe one possible way to implement a self-contained (server-side) and secure rate-limiting mechanism for web applications (browser-server interaction).
In an ideal situation, rate-limiting login attempts would be a waste of time and resources. In the ideal situation, all credential holders of your system would be using a password/pasphrase (or an equivalent) which is strong enough to make it impossible for anyone to guess a single credential, even if they could fire huge amounts of tries against your login mechanism.
If you used a “password” which has, say, 2^128 worth of guessing entropy strength, no one could throw enough login attempts to guess the right password. The adversary would be only wasting her time.
In reality, the situation is usually far from ideal. Passwords like “Password3” and so on are easy to crack offline, but such passwords, and especially without rate-limiting, are also feasible to crack online.
This is why arguably most online systems employing credentials should make sure they make the adversary’s job guessing a correct credential as hard as possible. Which can be read as as time consuming as possible.
It is difficult to act against adversary who tries a few “most common passwords” from many different sources against every account on the system, but we can focus especially on the “multiple source against single account” and “single source” threats.
If we assume the adversary can mount the login attempts from, say, thousands of unique IP addresses and we do not want to lock the target account (permanently), we must make sure it takes as long as possible to try all the candidate passwords the adversary holds. And we want to do this in such a manner which does not consume endlessly our login server resources (recall rate-limiting with
sleep() PHP function).
Unfortunately it is presumably impossible to delay the adversary while never (temporarily) locking out the actual account owner, too. But we can adopt same kind of per-source rate-limit and global rate-limit approach used in popa3d to the web environment, to make online password guessing as slow as possible.
We need to set a limit for authentication attempts tied to the source IP address and source IP address block. In general, the limit for the IP address block (say, 220.127.116.11 - 18.104.22.168) will be greater than the limit for a single IP address.
We could allow max. 25 attempts per single IP address and 100 attempts per
/24 IPv4 block within the last 10 seconds. Especially the 100 attempts limit against a /24 block is relatively pessimistic (assuming the adversary has such IP address block(s) in her control). Also, people behind NAT or sharing a proxy may encounter problems with the 25 attempts limit. Make sure you adjust these numbers for your situation.
IPv6 addresses are probably good to address in
Note you may want to disallow all (login) requests from you local private network.
We can also measure attempts against a specific account coming from multiple IP blocks (i.e. from a botnet spanning to hundreds or thousands computers around the world) and thus set per-account rate-limit.
For example, allow max. 5 distinct IP blocks to try to login against a single account within the last 10 seconds. The “5 distinct IP blocks” in this example is set to make it harder to harass the account owner by not letting him to login by sending bogus passwords for their username.
In addition to the per-source and per-account limits, we can measure attempts globally to make sure any attack (attack which isn’t already caught by the above rate-limits) can’t run faster than X we set.
We could allow max. 300 login attempts (no matter where they are coming from or what account they are targeting) within the last 10 seconds.
This sets the “guaranteed upper bound rate” an adversary can at most try the candidate passwords (per 10 seconds in this article). I.e. an attack coming from thousands of unique IP blocks and targeting several accounts can’t proceed faster than this global limit allows.
The database structure for the “login_log” table could look like this:
Note we use field
remote_ip_block to indicate the IP block the request came from. For example, if the request comes from address 22.214.171.124, we’ll store
inet_pton('126.96.36.199') (in PHP) as the IP block. This is used to query distinct IP blocks when calculating per-account rate-limit. With IPv6 This ought to be the same value as stored in the
First, when a login request arrives, we’ll check if it is allowed to attempts a login:
This gives us
If global_attempt_count is greater than 300 or
if ip_attempt_count is greater than 25 or
if ip_block_attempt_count is greater than 100 or
if username_attempt_count is greater than 5
reject the login.
This makes the adversary to wait (or change source addresses or target account).
Unfortunately it also affects the actual account holder or an user who is trying to login from the same source address or block where the attack is running. But as earlier said, this can’t be entirely avoided, so it’s good to tune the “login limiting parameters” as suitable for your use case as possible.
If the login attempt is not rejected proceed with the login.
If the credentials are correct, log the user in. If authentication fails, insert a new row into the log database.
Inserting a new row could go something like this (in PHP):
In addition to setting a “hard limit” for the current client, it is probably a good idea to add CAPTCHA and/or require 2-Factor Authentication after a certain threshold (server-side “hard limits” and client-side challenge-response tests are not mutually exclusive).
Besides CAPTCHA and 2FA, there are also other countermeasures to deal with online brute-force attacks which can be used in conjunction with rate-limiting, for example, see Device Cookies.
Regarding the above database, it may be good to delete older records periodically to make sure the log database doesn’t grow too big (and waste space and resources).
Also, if you consider in-memory databases, like Redis, to hold the logging information, pay attention to the “persistence problem” where older records are erased and how to query IP ranges. (Although needing to hold the records for 10 seconds may not be a problem).
The above rate-limiting methods sets upper bounds to different kind of attacks.
The most “basic attack” (coming from just one address) is limited to 25 attempts per 10 seconds.
Attacks coming from multiple hosts are limited to 100 attempts per 10 seconds when coming from a single /24 IPv4 block.
Attacks against a single username are limited to span at maximum to 5 different IP blocks per 10 seconds.
And finally, there is 300 attempts global limit (per 10 seconds).
The above described method is a self-contained (no external special hardware required) method which can be modified precisely for your situation to detect brute-force attempts, and act against these attacks. However, even aggressive login rate-limiting can not save poor passwords, so make sure no weak passwords enters in the system.
In the other side of the coin, the login related database handling will require resources. As for single login two database queries must be run (SELECT and INSERT). Although the INSERT only happens when the rate-limits are not reached.
After all, this involves once again traditional security vs usability trade-offs.
11 May 2015 This post was revised to address the critical problem with "login tokens", see this comment for more information. Sorry for the inconvenience!
comments powered by Disqus