Protecting websites and applications from cyber-threats is vital for any organization. A Web application firewall (WAF) prevents attacks to damaging applications. This provides a web security by filtering and monitoring traffic network to protect against attacks. A WAF solution based on the anomaly detection can identify zero-day attacks. Deep learning is the state-of-the-art method that is widely used to detect attacks in the anomaly-based WAF area. Although deep learning has demonstrated excellent results on anomaly detection tasks in web requests, there is trade-off between false-positive and missed-attack rates which is a key problem in WAF systems. On the other hand, anomaly detection methods suffer adjusting threshold-level to distinguish attack and normal traffic. In this paper, first we proposed a model based on Deep Support Vector Data Description (Deep SVDD), then we compare two feature extraction strategies, one-hot and bigram, on the raw requests. Second to overcome threshold challenges, we introduce a novel end-to-end algorithm Auto-Threshold Deep SVDD (ATDSVDD) to determine an appropriate threshold during the learning process. As a result we compare our model with other deep models on CSIC-2010 and ECML/PKDD-2007 datasets. Results show ATDSVDD on bigram feature data have better performance in terms of accuracy and generalization.