Although the dramatic increase in Online Social Network (OSN) usage, there are still a lot of security and privacy concerns. In such a scenario, it would be very beneficial to have a mechanism able to assign a risk score to each OSN user. For this reason, in this paper, we propose a risk assessment based on the idea that the more a user behavior diverges from what it can be considered as a `normal behavior', the more it should be considered risky. In doing this, we have taken into account that OSN population is really heterogeneous in observed behaviors. As such, it is not possible to define a unique standard behavioral model that fits all OSN users' behaviors. However, we expect that similar people tend to follow similar rules with the results of similar behavioral models. For this reason, we propose a risk assessment approach organized into two phases: similar users are first grouped together, then, for each identified group, we build one or more models for normal behavior. The carried out experiments on a real Facebook dataset show that the proposed model outperforms a simplified behavioral-based risk assessment where behavioral models are built over the whole OSN population, without a group identification phase.

The aim has been to propose a fraud detection system with capabilities of minimizing false alarms. In this paper we introduce a technique, which uses a hybrid fraud scoring and spike detection technique in streaming data over time and space. The technique itself differentiates normal, fraud and ano malous links, and increases the suspicion of fraud links with a dynamic global black list. Also, it mitigates the suspicion of normal links with a dynamic global white list. In addition, this technique uses spike detection technique to highlight the sudden and sharp rises in data, which can be indicative of abuse. The purpose is to derive two accurate suspicion scores for all incoming new examples in real-time. Results on mining several thousand credit application data demonstrate that the proposed technique reduces false alarm rates while maintaining a reasonable hit rate. In addition, new insights have been observed from the relationships between examples. The proposed technique takes the advantages of anomaly detection and supervised techniques. However by employing the spike detection technique, the false alarm rate is decreased. By this novel integration of techniques, the proposed technique is able to foil fraudsters' attempts, which continuously morph their styles to avoid to be detected. The results of the experiments to demonstrate the benefits of the technique are also presented in this paper.