Low-rate Denial of Service (LDoS) attacks use the low-rate requests to achieve the occupation of the network resources and have strong concealment. The traditional signal analysis based detection methods are challenging to detect LDoS attacks in the fluctuating legitimate traffic. In this paper, an LDoS attack detection method based on hybrid deep neural networks is proposed using one-dimensional convolutional neural network and gated recurrent unit. In order to evaluate the proposed detection method in the real scenarios, we captured real legitimate traffic from a website in the datacenter, and carried out a variety of real LDoS attacks on the mirror of the website in the laboratory environment to obtain real attack traffic. The detection results on the real traffic show that the proposed detection method does not need to extract features manually and can effectively detect LDoS attacks in fluctuating HTTP traffic with an average detection rate of 98.68%, which is more advantageous than MF-DFA or power spectral density based detection methods.