Hierarchical Knowledge Distillation and Pruning (HKDP) Framework for Secure and Sustainable Federated Learning

The Hierarchical Knowledge Distillation and Pruning (HKDP) framework project we propose addresses the growing demand for real-time data processing in cybersecurity applications. Traditional centralized machine learning approaches are hindered by high computational requirements and privacy concerns, but federated learning (FL) offers a solution by enabling decentralized model training, maintaining data privacy, reducing energy consumption, privacy, and security assurance.
This research project aims to design, implement, and evaluate the HKDP framework for optimizing energy efficiency in FL for cybersecurity. The objective is to enhance real-time data processing while minimizing computational burdens on devices. To achieve this, the project will pursue four key sub-objectives: conducting a thorough literature review, designing the framework, implementing the HKDP framework, and experimentally evaluating its performance.
The project will investigate the impact of hierarchical knowledge distillation on energy efficiency, optimal pruning strategies, and the adaptability of energy-aware FL frameworks. Utilizing established cybersecurity datasets and PyTorch/TensorFlow implementation, evaluation metrics will include energy consumption, model accuracy, and computational requirements.
The anticipated outcomes include improved energy efficiency (up to 30% reduction), enhanced model accuracy (up to 5% improvement), and reduced computational requirements (up to 20% decrease). The HKDP framework will provide an innovative solution for energy-efficient federated learning in cybersecurity.

Faculty Supervisor:

Haïfa Nakouri

Student:

Partner:

Instituto Tecnológico Autónomo de México;National School of Advanced Technologies Borj Cedria - Tunisia

Discipline:

Computer science

Sector:

Artificial Intelligence; Cyber Security; Sustainability and the Environment

University:

Université du Québec à Chicoutimi

Program: