Cyber security attacks has always been a big threat to any known smart and connected industry, not only the economical losses are great but also in term of data privacy, this is why it is essential to put in place preventive measures capable to detect those malicious attacks as reliable and as early as possible. For this matter anomaly detection comes handy as a technique of early prevention by stopping those attacks through a large set of methodologies such us machine learning and deep learning models. for this work we are going to focus on the elaboration a benchmarks of different solutions and focus on a new recent approach of detection using a Graph neural networks applied to a Bluetooth dataset, that contains many attacks such us the DOS and the DDOS and the Bluesmack. We will begin by analyzing the dataset through a set of methodologies starting with a baseline analysis and machine learning techniques: Isolation forest, support vector machines and continue with deep learning models: LSTM autoencoders and graph convolutional networks. The work is going to outline the different aspects of anomaly detection, a walk-through of the dataset with the different components of the Bluetooth stack layer, the injected attacks and the impact on the dataset, then we’ll define each approach of machine learning and deep learning techniques,moving to an experimental section of each method and their evaluation performances. Finally a discussion of the potential improvements of those techniques and how it can be extended or optimized afterward.