Millions of dollars are lost due to fraud every year and there is a growing realization in organizations that a significant amount of fraud is committed from inside a corporation. Previous fraud detection systems used deep learning and manual feature engineering without taking relations between different entities into account. With the emergence of the geometric deep learning models we are able to process features related to an entity while considering the relationships between these entities. GCN models, one of the more popular geometric deep learning models, have already proven their success in fraud detection. However, the models used in previous research make use of external features to detect anomalous activities, such as topic and sentiment analysis of the content of the emails, information which is hardly ever available in a real-world organization. Therefore, in this paper, we develop a methodology to extract features linked to structural anomalies and apply them in a GCN model. Our proposed model is evaluated on the Enron Email Corpus, a real-world insider threat dataset.