Man-edos: A multihead attention network for the detection of economic denial of sustainability attacks

Abstract

Cloud computing is one of the most modernized technology for the modern world. Along with the developments in the cloud infrastructure comes the risk of attacks that exploit the cloud services to exhaust the usage-based resources. A new type of general denial attack, called “economic denial of sustainability” (EDoS), exploits the pay-per-use service to scale-up resource usage normally and gradually over time, finally bankrupting a service provider. The stealthiness of EDoS has made it challenging to detect by most traditional mechanisms for the detection of denial-of-service attacks. Although some recent research has shown that multivariate time recurrent models, such as recurrent neural networks (RNN) and long short-term memory (LSTM), are effective for EDoS detection, they have some limitations, such as a long processing time and information loss. Therefore, an efficient EDoS detection scheme is proposed, which utilizes an attention technique. The proposed attention technique mimics cognitive attention, which enhances the critical features of the input data and fades out the rest. This reduces the feature selection processing time by calculating the query, key and value scores for the network packets. During the EDoS attack, the values of network features change over time. The proposed scheme inspects the changes of the attention scores between packets and between features, which can help the classification modules distinguish the attack flows from network flows. On another hand, our proposal scheme speeds up the processing time for the detection system in the cloud. This advantage benefits the detection process, but the risk of the EDoS is serious as long as the detection time is delayed. Comprehensive experiments showed that the proposed scheme can enhance the detection accuracy by 98%, and the computational speed is 60% faster compared to previous techniques on the available datasets, such as KDD, CICIDS, and a dataset that emerged from the testbed. Our proposed work is not only beneficial to the detection system in cloud computing, but can also be enlarged to be better with higher quality of training and technologies.