This scan type works by exploiting "predictable IP fragmentation ID" sequence generation on the zombie host, to determine open ports on the target. The scan checks the IPID on the zombie, then spoofs a connection request to the target machine, making it appear to come from the zombie. If the target port is open, a SYN/ACK session acknowledgement will be sent from the target machine back to the zombie, which will RST the connection since it has no record of having opened such a connection. If the port on the target is closed, an RST will be sent to the zombie, and no further packets will be sent. The attacker then checks the IPID on the zombie again. If it has incremented by 2 (or changed by two steps in its sequence), this corresponds to the packet received from the target, plus the RST from the zombie, which equates to an open port on the target. If the IPID has changed by one step, an RST was received from the target and no further packets were sent.
Using this mechanism, it is possible to scan every port on a target, whilst making it appear that the zombie was the one doing the scanning. Of course, the spoofed connection attempts will likely be logged, so the target system will have the zombie IP address, and the zombie system's logs are likely to contain the attacker's IP address, so it is still possible, after acquiring logs through legal channels, to determine the attacker, but this method makes it much more difficult to do so than if the packets were sent directly from the attacker. In addition, some IDS and firewall software makes attempts to detect spoofed packets based on the network they arrive from. As long as the zombie host and the attacker are both "out on the Internet", or on the same network as each other, relative to the target, techniques to identify spoofed packets are not likely to succeed.
This scan type requires certain things of the zombie. The IPID sequence generation must be predictable (single-step increments, for example). The host must also have low traffic so that it is unlikely for other packets to hit the zombie whilst Nmap is carrying out its scan (as these will artificially inflate the IPID number!). Cheap routers or MS Windows boxes make good zombie hosts. Most operating systems use randomised sequence numbers (see the OS Fingerprinting section for details on how to check a target's sequence generation type).
The idle scan can also be used to determine IP trust based relationships between hosts (e.g. a firewall may allow a certain host to connect to port x, but not other hosts). This scan type can help to determine which hosts have access to such a system.
For more information about this scan type, read http://www.insecure.org/nmap/idlescan.html