A new multistage approach to detect subtle DDoS attacks

Detection of distributed denial of service (DDoS) attacks has been a challenging problem for network security. Most of the existing works take into account the anomaly features of the traffic caused by DDoS. However, these detection methods suffer from either less generality or high computational and memory costs in detecting subtle DDoS attacks. In this paper, we first present a model for DDoS attacks with quantitative measurements. Based on this model, we find that there are two factors that have a severe influence on the deviation of traffic features. In view of these two factors, the DDoS attack traffic observed by monitors can be trivial, leading to the subtle DDoS attacks which are difficult to detect. To detect the subtle DDoS anomalies at monitors close to the attack sources, we propose a novel multistage DDoS detection framework that consists of a NTS (Network Traffic State) prediction, a fine-grained singularity detection and a malicious address extraction engine. We also briefly introduced how to distribute our detection framework to enhance the performance of detecting world-wide DDoS attacks. Moreover, the prototype system is implemented and evaluated with real network traces from our campus network and testbed. The results show that our method can detect various DDoS attacks efficiently even though the attack rate is low. Our method can extract malicious IPs for attack reaction with records for a short period, and multiple monitors distributed in the network can fuse the results of extraction seamlessly to improve the accuracy of detection