Srikanth Kandula, Dina Katabi, Matthias Jacob, and Arthur Berger
Recent denial of service attacks are mounted by professionals using Botnets of tens of thousands of compromised machines. To circumvent detection, attackers are increasingly moving away from bandwidth floods to attacks that mimic the Web browsing behavior of a large number of clients, and target expensive higher-layer resources such as CPU, database and disk bandwidth. The resulting attacks are hard to defend against using standard techniques, as the malicious requests differ from the legitimate ones in intent but not in content.
We present the design and implementation of Kill-Bots, a kernel extension to protect Web servers against DDoS attacks that masquerade as flash crowds. Kill-Bots provides authentication using graphical tests but is different from other systems that use graphical tests. First, Kill-Bots uses an intermediate stage to identify the IP addresses that ignore the test, and persistently bombard the server with requests despite repeated failures at solving the tests. These machines are bots because their intent is to congest the server. Once these machines are identified, Kill-Bots blocks their requests, turns the graphical tests off, and allows access to legitimate users who are unable or unwilling to solve graphical tests. Second, Kill-Bots sends a test and checks the client's answer without allowing unauthenticated clients access to sockets, TCBs, and worker processes. Thus, it protects the authentication mechanism from being DDoSed. Third, Kill-Bots combines authentication with admission control. As a result, it improves performance, regardless of whether the server overload is caused by DDoS or a true Flash Crowd.
In Second Networked Systems Design and Implementation (NSDI)
All copyrights reserved by USENIX 2007