full-disclosure - unix entropy source can be used for keystroke timing attacks

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [day] [month] [year] [list]

Message-ID: <20030815010139.E77480-100000@dekadens.ghettot.org>
From: lcamtuf at ghettot.org (Michal Zalewski)
Subject: unix entropy source can be used for keystroke timing attacks

Another bizarre vulnerability, for your amusement...

Several unix systems systems provide a secure entropy source maintained by
collecting certain information that is supposed to be practically
unpredictable (such as interrupt timings, keyboard scancodes or device
request times), then running it thru one-way hashing function (SHA1 or
MD5).

A device called /dev/random estimates the amount of entropy available in
the entropy pool, and blocks on read when the pool gets empty. This
continues until the buffer is seeded with some new data due to one of the
events mentioned.

By emptying the /dev/random, and then timing the moments when data becomes
available, we can precisely determine, at what intervals those events
occur. While we can't determine what data is being added to the entropy
pool, we can rather easily tell a situation when a keystroke data is added
due to a specific pattern triggered by it. For example, on Linux the
sequence is:

  Keypress scancode in: one to two bytes available
  50-150 ms delay (keypress duration)
  Key release scancode in: one to two bytes available
  50 ms or more delay

Other types of events, such as disk activity, usually generate a burst of
events, usually under 1 ms away from each other, or have other distinct
patterns (besides, those events happen only sporadically).

Because of this, it is possible to measure keypress AND key release
timings _very precisely_, for any console user of a machine we have an
unprivileged account on. Timings between keystrokes depend on the distance
between subsequently pressed keys for each hand, and the placement of the
hand, keystroke durations usually depend on the finger used - all this
making it quite easy to come up with a nice subset of possible passwords,
and not impossible to determine some of the commands typed.

There is some mature research in the field of recovering typed information
or its certain properties from the timing information (both for the
purpose of biometrics and surveillance), so I don't think to get too far
into this. I can just provide an example of the difference in average
keystroke timings for two related words typed by a keyboard-proficient
user:

  1) "evil"

        e press      |
        e release    | =========
        v press      | ===============
        i press      | ====
        v release    | ==
        i release    | ====
        l press      | =======
        l release    | ============

        Note out of sync 'v' release after hand switch.

  2) "good"

        g press      |
        g release    | ========
        o press      | ====
        o release    | =====
        o press      | =======
        d press      | =========
        o release    | =======
        d release    | ==============

        Same for second 'o'.

A workaround would be to add some latency before unblocking pending
read()s when a new information becomes available. /dev/random is a very
high latency device, and no program depends on how fast the data is
available for its normal operation.

-- 
------------------------- bash$ :(){ :|:&};: --
 Michal Zalewski * [http://lcamtuf.coredump.cx]
    Did you know that clones never use mirrors?
--------------------------- 2003-08-15 01:01 --