Banca de QUALIFICAÇÃO: OGOBUCHI DANIEL OKEY

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
DISCENTE : OGOBUCHI DANIEL OKEY
DATA : 06/12/2024
HORA: 14:00
LOCAL: https://conferenciaweb.rnp.br/sala/joao-74
TÍTULO:

Enhancing the Privacy and Security of Intrusion Detection Systems in IoT through Federated Learning

PÁGINAS: 81
RESUMO:

Information transmission in the Internet of Things (IoT) era, which has highly infiltrated virtually all facets of human endeavours, has become an issue of grave importance. The rate of increase in vulnerabilities makes these systems susceptible to diverse threats, that are exploited by cybercriminals to cause mayhem, hence, necessitating urgent responses through advanced privacy and security measures. Machine learning (ML) and deep learning (DL) use centralized mechanism to train algorithms for IoT security. Despite the performance of the centralized learning (CL) process, major concerns such as privacy, data ownership, and high computational costs still exist. Federated learning (FL) applies advanced approach to address the weaknesses of CL, which holds sensitive IoT data for model development in a particular system by maintaining the privacy of client data, providing more secure data transmission, reducing network bandwidth, and providing flexibility in data usage. However, several challenges, including communication efficiency, data privacy, high model parameters, real-time response, and scalability of FL, still pose hindrances towards the deployment of FL-based intrusion detection system (IDS) models in resource constrained IoT devices. The objective of this proposal is to develop an IDS solution that is optimized for both privacy and network communication efficiency in the IoT context. Our proposed solution involves enhancing the algorithm and model through compression and quantization techniques, which would address the communication overhead cost. Knowledge distillation is adopted as a model compression method. In addition, we prioritize privacy-centric measures, such as differential privacy and secure multiparty computation in the algorithm design to ensure that model parameter exchange is secure to a high level. Initially, we develop model based on convolutional neural network (CNN), gated recurrent network (GRU) and Long Short Term memory (LSTM) and achieved an accuracy of about 95%. Following our understating that there are too many feature space in the training data, we developed a new algorithm to reduce data diemnsionality, with an accuracy of 98% using top 20 features. Furthermore, we investigated KD to reduce the model parameters with varying accuracy on the datasets. Based on the results, we ultimately aim to offer a comprehensive and robust solution to the privacy and communication challenges that arise in FL for IoT networks.

MEMBROS DA BANCA:
Presidente - Interno ao Programa - 1603840 - JOAO HENRIQUE KLEINSCHMIDT
Membro Titular - Examinador(a) Interno ao Programa - 1761107 - RICARDO SUYAMA
Membro Titular - Examinador(a) Externo ao Programa - 1673092 - RONALDO CRISTIANO PRATI
Membro Suplente - Examinador(a) Interno ao Programa - 2196309 - CARLOS ALBERTO KAMIENSKI