مرتضي فائزي نيا

Governments and companies are generating massive amounts of data over time. Taking advantage of these data requires efficient analysis and extraction of valuable information. Patterns in such data change over time and reduce the efficiency of existing models in learning systems. To face this challenge, the current models should have adaptability over time. and adapt to the change of statistical characteristics and the emergence of new patterns and prevent the emergence of predictions and poor results for decision-making. An essential part of these analyzes is anomaly detection. Identifying abnormal behaviors in systems and processes prevents damage, and intrusions and takes appropriate actions in less time. In common approaches, model learning is usually offline and uses past data. This happens in dynamic and unstable conditions where the definition of natural behavior changes over time and reduces the effectiveness of the existing model. In the problem of anomaly detection, this thesis proposes solutions for the three challenges of dynamic data, robustness against noise, and adapting to pattern change using deep learning methods and considering dynamic data processing. These solutions are adapted by changing data behavior and have the ability to detect and explicitly adapt to anomalies. The proposed approaches have been investigated and eva‎luated with the real data set and their performance in terms of anomaly detection power, convergence, and adaptation of the learner model against new data compared to similar methods according to the F1 criterion for the first and second proposed approaches respectively, 3 and 12 percent have improved.

تشخيص ناهنجاري , تغيير الگو , پويايي‌داده , يادگيري ژرف

Anomaly detection , Concept Drift , Dynamic Data , Deep Learning

Morteza Faezinia

Dr. Behrooz Minaei Bidgoli