Ανίχνευση εισβολών με τη χρήση βαθιάς και μη κεντρικοποιημένης μάθησης

Abstract

In the frantic momentum that today’s technology is rapidly progressing, both people, atomically, and organizations are trying to keep up with the latest and safest ways to procure their everyday interactions with the digital world. More often than not, though, malicious entities, those being socioeconomical or deterministic elements, try to undermine the normal way that systems work and to actively exploit users or companies by overcoming the security mechanisms that protect their privacy and function. To tackle this problem and ensure the protection of the privacy and safe operation of digital services, the field of cybersecurity strives to minimize the invasion of malicious entities into personal or commercial networks. One of the implemented techniques used in the field of cybersecurity are the Intrusion Detection Systems (IDS) that stand to recognize and in extend prevent an attack to protect digital resources. By using already known digital system interaction and through the methods of Machine Learning and Deep Learning, build and train models that detect such attacks. A problem that arises, though, with the anonymization of the data used to train IDS systems. Federation Learning, a novel decentralized method of training and communicating deep learning models addresses this problem by decoupling the data from the training process of centralized systems. This work delves into the building of Intrusion Detection systems with deep learning algorithms and adapting them to the powerful collective process of Federation Learning.