A link-flooding DDoS attack works by congesting a few chosen network links that carry traffic toward a target, rather than by sending anything to the target itself. The target's servers stay quiet while its users are cut off, which makes it one of the few attacks a DDoS test cannot reproduce on the real network, and one where the test's attention moves from the target to the paths around it. The best-known form is the Crossfire attack, published by Kang, Lee and Gligor at the IEEE Symposium on Security and Privacy in 2013.
How the Crossfire attack works
A network region is reached through a limited number of links, and many routes into it share the same few. Crossfire picks those shared links as the real target.
The attacker's botnet then sends flows not to the target but to public servers near it, the decoys. Each flow is low-rate and ordinary, a connection to a server that would accept it from anyone. On their way to the decoys, the flows meet on the chosen links, and together they fill them. Traffic for the target crosses the same links and is lost with everything else.
Three properties follow:
- The target receives nothing. No packet is addressed to it, so nothing reaches its servers or its mitigation.
- No destination stands out. The load is spread across many decoys, each seeing modest traffic it would normally serve.
- The flows look ordinary to routers. Every flow is low-rate and well formed, from a real source, so per-flow inspection finds nothing to drop. Only the total on the shared link is abnormal.
That is what separates it from a plain volumetric flood, where the flood is aimed at your address and the size of it is the evidence.
Why target-side defence cannot see it
Every defence that works on traffic addressed to you is out of the path. Diverting your prefixes to a scrubbing service gives it nothing to filter, because the flood is addressed to someone else. A clean pipe filters traffic bound for your prefixes on your carrier's network, and none of this traffic is bound for them. Rate limits, challenges and detection on your own edge see a drop in real users and no attacker at all.
The symptom is also unusual. A flood normally arrives as a spike on your own graphs. Here the inbound graph falls, and the only clear sign is that users on certain networks report they cannot connect.
So resilience comes from the shape of the network around you, not from a filter. It depends on how many genuinely separate upstream paths you have, on whether they share links further out, and on how quickly traffic can be moved onto the ones still working, which is the job of DDoS failover.
What a DDoS test can measure
A test may not flood links that belong to third parties, and it may not use other people's servers as decoys. No one can authorise that, so a link-flooding attack is never reproduced as such.
What a test can measure is the exposure:
- Upstream diversity. Map the paths toward your prefixes from many networks and see where they converge. Two carriers that both hand off through the same upstream link are one path with two contracts.
- Failover under load. With your own carriers in the plan, take one upstream out of service during the window and time how long until traffic flows through the rest, and whether the remaining capacity holds normal peak plus a test flood.
- Detection from outside. Inbound telemetry shows a drop but not why; synthetic checks from several external networks show which paths have failed. A test confirms they alert, and how fast.
Whether a second path actually carries the service once the first is gone, measured with load on it rather than assumed from a diagram, is the subject of DDoS failover testing.