Loading...
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 55824 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Amini, Morteza
- Abstract:
- With the growing trend of IoT, especially in critical areas like health system and city management, and the expectations of even higher growth with the advent of 5G networks, the security and preserving of users' privacy in IoT has gained significant importance. Anomaly detection is one of the approaches to monitor IoT devices which enables the identification of anomalous behaviors. This anomalous behavior could indicate malware infection, physical malfunctions, or tampering.Deep learning has been a common approach for anomaly detection for the past few years. The solutions are mostly suggested in a special purpose manner and because they are based on a particular deep learning model, they are practically incapable to operate in diverse environments. Furthermore, these solutions are based on central computing and all of the training and inference from the models is carried out in a single central cloud-server. The fact that these systems heavily depend on the internet infrastructure and lack fault-tolerance results in less availability and a higher chance of failure in the whole anomaly detection system. This research proposes a flexible, comprehensive solution for anomaly detection in IoT environments upon a suggested framework. This framework is constructed by introduction of micro-model concept, distribution of models training and inference tasks in various levels, and use of cloud-edge architecture. In this architecture, the edges are responsible for training and inference of anomaly and malware detection micro-models, and the servers coordinate micro-models training and aggregation procedures. The leader server handles coordination and distribution of training and aggregation load of edge micro-models over different servers. Selection of the leader server and synchronizing status data among the servers is implemented with Raft consensus algorithm. These characteristics help avoiding single central server dependency, utilize computing resources, preserve edge nodes privacy, and reduce connection overhead between servers and edges. This reduction in connection overhead which stems from transferring micro-models instead of data, makes it possible to use the suggested anomaly detection framework in practice for diverse IoT environments. To assess the proposed solution, a cloud-edge architecture in IoT environment is simulated using virtual machines. The simulation environment consists of three devices and eight popular IoT malwares. The performance of the micro-models is evaluated based on multiple criteria. The results for general performance evaluation of micro-models in terms of accuracy, precision, recall, and sensitivity are 99.77, 99.77, 99,96, and 99.96, respectively, which indicates great performance of the suggested solution
- Keywords:
- Internet of Things ; Anomaly Detection ; Distributed Learning ; Edge Computing ; Cloud-based Architecture ; Cloud Security
Digital Object List
-
محتواي کتاب
- view
Bookmark