With the rapid development of Internet and its services, cyber attacks are increasingly emerging and evolving nowadays. To be aware of new attacks and elaborate the appropriate protection mechanisms, an interesting idea is to attract attackers, then to automatically monitor their activities and analyze their behaviors. In this paper, we are particularly interested in detecting and learning attacks against web services. We propose an approach that describes the attacker’s behavior based on data collected from the deployment of a web service honeypot. The strengths of our approach are that (1) it offers a high interaction environment, able to collect valuable information about malicious activities; (2) our solution preprocesses the set of data attributes in order to keep only significant ones (3) it ensures two levels of clustering in order to produce more concise attack scenarios. In order to achieve these contributions, we employ three analysis techniques: Principal Component Analysis, Spectral Clustering and Sequence Clustering. Our experimental tests allow us discovering some attacks scenarios, such as SQL Injection and Denial of Services (DoS), that are modeled in Markov chains.