KAIST researchers demonstrate remote AI model extraction via electromagnetic signals
Researchers at KAIST reported developing ModelSpy, a system that can remotely extract AI model structures by capturing electromagnetic signals from GPUs via a small antenna at distances up to 6 meters, achieving up to 97.6% accuracy. The research demonstrates a potential security vulnerability for enterprise AI systems and proposes countermeasures including electromagnetic shielding and computation obfuscation.
What happened
KAIST researchers led by Professor Han Jun developed a system called ModelSpy that can remotely extract AI model structures using a small antenna.
The system captures faint electromagnetic signals emitted by graphics processing units (GPUs) when performing computations, then analyzes these signals to infer the AI model architecture.
In experiments on five latest-generation GPUs, researchers achieved up to 97.6% accuracy in estimating model structures from beyond walls and at distances up to 6 meters.
Reported impact
The research indicates that critical AI assets in enterprises could potentially be exposed to external theft without direct server intrusion, representing a new security threat.
The researchers emphasized that important AI infrastructure, such as autonomous vehicles and national infrastructure systems, requires protection.
Index analysis
The attack targets the physical side-channel emanations from GPUs during AI computations, exploiting unintended electromagnetic radiation.
The research was presented at NDSS 2026 and received the best paper award.
Governance controls
The researchers proposed countermeasures including electromagnetic shielding and computation obfuscation techniques.
Researchers called for establishing integrated cyber-physical security systems combining hardware and software protections to safeguard critical AI infrastructure.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.