Every cloud server runs a hypervisor, the software layer that lets many virtual machines share one computer. If it is hacked, everything on that server is exposed. This research uses federated learning to build defences that adapt as attackers change tactics.
The problem
A static security layer cannot keep up with evolving cyberattacks on hypervisors, such as hyperjacking, side-channel attacks and virtual machine escape. Security needs to adapt as attackers do.
What the research did
We designed an adaptive communication protocol enhanced by federated learning. Federated learning lets AI models on many hypervisors learn together without sharing their raw data, which protects privacy.
Local and global anomaly detection models work together to adjust security protocols dynamically when threats appear.
What we found
The adaptive approach detected attacks with 92.6% accuracy, compared with 85.2% for centralised learning and 78.4% for static protocols. It also reduced communication overhead by 55% and training time by 32%.
Who benefits
- Cloud providers protecting shared infrastructure
- Businesses that run critical services in the cloud
- Privacy-sensitive organisations, since data stays local
Limits and what comes next
The results come from experiments; real cloud environments are larger and more varied. Federated learning can itself be targeted by poisoned updates, which future work needs to address.
The team
This research was led by Moutaz Alazab and Albara W. Awajan (Al-Balqa Applied University) with Areej Obeidat, me, Aaron Bere (Torrens University Australia), Saqib Ali (Sultan Qaboos University) and Wei Wei (Xi’an University of Technology).
Frequently asked questions
What is a hypervisor?
Software that lets one physical server run many virtual machines. It is a high-value target for attackers.
What is federated learning?
A way for AI models in many places to learn together without sending their raw data to one central location.
How accurate is this approach?
92.6% detection accuracy, higher than centralised (85.2%) and static (78.4%) approaches.