TCP SYN attacks - a simple solution

Matthew Kaufman matthew at scruz.net
Sun Oct 6 23:30:14 UTC 1996


Original message <199610062314.TAA29781 at merit.edu>
From: rex at staff.cs.su.oz.au (Rex di Bona)
Date: Oct  7,  8:10
Subject: TCP SYN attacks - a simple solution
> 
...
> I propose a solution where the initial sequence number is calculated
> (not random), and is based on a cryptographic calculation of the
> senders Initial Sequence Number, the ports, and a "per boot"
> secret number. In this way the initial packet can be discarded,
> and on receipt of the third SYN packet can be recalculated.
...

The idea has been floated before, and I believe it to be the right
solution to this problem. However, I have some suggested improvements:

1. The use of a "per boot" secret number allows an attacker to
   poll your machine to deduce the secret, and then attack you with
   that knowledge. 

   A solution to this problem is to use a rapidly changing secret, the
   pattern of which cannot be easily deduced, and a sliding window of
   acceptance. (If the hash doesn't match the current scheme, but matches
   the scheme we were using in the past N seconds, then accept the packet)

   The change interval needs to be short enough that, by the time an
   attacker has been able to compute the next number, the window for
   accepting that has closed.

2. The TCP specification allows data to ride along with the initial SYN,
   and requires that after the three-way handshake, that the data be
   presented to the application layer.

   One solution is to realize that very few implementations take advantage of
   this, and simply not accept this form of SYN.

   A second solution is to NOT ack the data that is riding along, but the
   TCP state machine definition has some holes here which may not work for
   all implementations.

   The solution of buffering those SYN's which contain data is unacceptable,
   because attackers will simply switch to sending SYNs with data.

I believe that switching to this sort of scheme would also thwart most
attacks which rely on sequence number prediction, and save memory significantly
over schemes which simply increase the number of TCBs allowed.

-matthew kaufman
 matthew at scruz.net

ps. I've been meaning to write this entire scheme, with the enhancements
  I propose here, as a draft specification, but I keep getting interrupted
  by flooded phone rooms and the like this weekend. *sigh*






More information about the NANOG mailing list