Some cyberattacks start with someone physically walking up to a server room or power facility. This research uses AI to watch security camera footage, recognise suspicious behaviour around critical sites, and raise the alarm with minimal human effort and very low cost.
The problem
Cyber-physical attacks damage both digital and physical assets, causing data theft, financial loss and infrastructure damage. Defending against them is harder than pure software security because it needs people to physically watch and inspect sites, around the clock.
What the research did
We built an algorithm that recognises human activities in real-time CCTV footage around the perimeter of critical infrastructure. It combines GoogleNet, a network that recognises what is in each video frame, with a bidirectional LSTM that understands how movements unfold over time.
To keep it responsive even on slow networks, it uses Machine Vision at the IoT Edge (Mez) technology. The AI runs from a cloud server, which adds a second layer of protection for the security system itself.
What we found
The system classified suspicious activities from live video with an average accuracy of 73.15%. Thanks to the optimisation scheme, it could strengthen cyber-physical security for an estimated $4.29 per month.
Who benefits
- Operators of data centres, utilities and telecom sites
- Organisations that cannot afford 24/7 human guards at every location
- Security teams who want automated alerts to supplement human monitoring
Limits and what comes next
An accuracy of about 73% means the system would miss or misread some events, so it is best used to support human security staff rather than replace them. Recognising human activities fairly across different people and settings also needs careful testing.
The team
This research was led by Sandesh Achar (Calix) with me, Md Whaiduzzaman (QUT), Albara W. Awajan and Moutaz Alazab (Al-Balqa Applied University).
Frequently asked questions
What is a cyber-physical attack?
An attack that affects both digital systems and physical equipment, for example someone gaining physical access to a server room to steal data or sabotage machines.
How does AI recognise suspicious behaviour?
It learns from labelled video examples to tell suspicious actions apart from normal ones, frame by frame and over time.
Does this replace security guards?
No. It is designed to help guards by flagging events they might miss.