On the 8th of June, the University College of Dublin, UCD, SPATIAL partner and represented by Dr Madhusanka Liyanage presented two accepted papers at the EuCnC in Grenoble, France.
Madhusanka is the Director of CySec Consulting, an IEEEr8 ComSoc Outstanding Young Researcher, one of the World’s Top 2% Scientists and a Marie Curie Fellow.
The papers presented were:
Paper1: Yushan Sririwardhane, Pawani Porambage, Madhusanka Liyanage, Mika Ylianttila, “Robust and Resilient Federated Learning for Securing Future Networks” in Proc. of 2022 Joint European Conference on Networks and Communications (EuCNC) & 6G Summit, Grenoble, France, June 2022.
“Robust and Resilient Federated Learning for Securing Future Networks”
In the past, machine learning has been more centralized because sufficient computing power was available only in centralized servers. Since the data needs to be sent to the centralized servers, there were issues with data privacy and high data transfer over the network. However, with the advancement of technology, sufficient computing power is now available at the end devices such as mobile phones to train machine learning models. With this, Federated Learning (FL) has become popular, which is a method of training a machine learning algorithm with the participation of a large number of distributed users (ex: mobile phones) coordinated by a central server. In FL, the data is kept in the devices, therefore it ensures privacy, and it does not utilize high network bandwidth.
Since the data is unavailable to the central server that coordinates all the users involved in the FL process, attacks called poisoning attacks are possible in FL systems. Simply, a poisoning attack manipulates the data and trains the model to provide erroneous outcomes to attacker’s chosen inputs. Therefore, the central server must have a mechanism to separate the attackers from the real users and remove the attackers from the system. These are called robust algorithms. The research exploits a vulnerability of an existing robust algorithm called FoolsGold, performs a coordinated poisoning attack to break its defence. In addition, the research also proposes a novel mechanism to identify such coordinated attacks, so that the central server can identify the coordinated attackers and remove them from the FL process.
Paper 2: Suwani Jayasinghe, Yushan Sririwardhane, Pawani Porambage, Madhusanka Liyanage, Mika Ylianttila, “Federated Learning based Anomaly Detection as an Enabler for Securing Network and Service Management Automation in Beyond 5G Networks” in Proc. of 2022 Joint European Conference on Networks and Communications (EuCNC) & 6G Summit, Grenoble, France, June 2022.
Abstract: The ZSM or Zero-touch Network and Service Management is an architecture designed for automating network operations. It incorporates machine learning/artificial intelligence for its operations, including securing the network. However, there are concerns such as data privacy and resource constraints when using machine learning for its operations. As a privacy-preserving technique, federated learning can be used. In FL, the model is generated in local devices, and data is not collected in a data center. As data is not collected in a data center it is also possible to ensure communication efficiency.
The proposed model inspects network flows using two detectors. The model in both these detectors is developed based on federated learning-based techniques. The first detector, which has a simpler structure than the second detector, inspects the network flows first. If the network anomalies are confirmed to be anomalies, those will be dropped. The network flows confirmed to be good are sent to the second stage. The second stage detector inspects the network flows again, and the network flows confirmed to be normal will be sent to their destination. As per the simulation, 93.5% is the highest accuracy achieved from this method. Also, an improvement is shown in all cases tested.