LAUSR

Cloud computing is now known as the most cost-effective platform for delivering big data and artificial intelligence services over the Internet to enterprises and cloud consumers. However, despite many recent security developments, many cloud consumers continue to express great concern about using these platforms because they still have significant vulnerabilities. Typically, Economic Denial of Sustainability (EDoS) attacks exploit the pay-as-you-go billing mechanisms used by cloud service providers, so that a cloud customer is forced to pay an extra fee for the additional resources triggered by the attack activities. In our previous work, we already proposed an system to mitigate such EDoS attacks. Overall, this previous work presented an effective system for detecting abnormal events; however, the false-alarm rates still remain relatively high and detection rates are low, because abnormal events could be caused by the cloud customer. Furthermore, our previous work still consumes a large number of computing resources. Therefore, in this paper, we propose an enhanced scheme to detect and mitigate EDoS attacks efficiently and reliably. Our proposed scheme is composed of online and offline phases, implementing a gated recurrent unit, which not only can capture complex temporal dependence relations in the data, but also can reduce the vanishing gradient problems in time series. First, to reflect the normal patterns, our proposed scheme learns accurate representations of multivariate time series. Next, these representations are used to reconstruct input data. Finally, the reconstruction probabilities not only can be used to find anomalies, but also can provide interpretations. The proposed scheme also introduces a self-adjusting threshold to reduce error rates, whereas existing solutions normally use a hard threshold to analyze the anomalies, which causes increasing error rates. Our comprehensive analysis of the results shows outstanding performance compared to other solutions and our previous work.