It is not just games and summers of the century that make processors break out in a sweat. Hacker attacks have a way of doing this as well. Temperature sensors combined with machine learning are being handed the job of spotting such threats at an early stage in the future.
“Major disaster” is the most frequently used expression to describe the announcement about serious security flaws in Intel processors roughly three weeks ago. Attackers can read protected data by slipping through gateways called Meltdown and Spectre. Intel processors are the primary target here. But AMD and ARM processors do not escape unscathed. This means one thing: Nearly all desktop computers, laptops, mobile devices and cloud services are vulnerable. A “major disaster” if there ever were one.
Data theft is just one possible consequence of cyberattacks. Another is sabotage. This form of destruction really has nothing to do with Meltdown and Spectre. As a result of the miniaturization efforts performed in recent years, it only takes “a few” electrons to run a program. In processors made in 10-nanometer production, even the slightest overload can trigger artificial aging processes that can result in the premature death of a chip. Such an event could be triggered by false control commands issued by a hacker. To defend against such attacks on places like industrial facilities, researchers at the Karlsruhe Institute of Technology (KIT) are working on a smart self-monitoring system.
Infrared cameras against cyberattacks
This system is based on the fact that every chip has a specific thermal fingerprint during normal operations. As a result, operations like calculations and access to memory or the hard drive cause brief periods of heating and cooling in various areas of the processor. The researchers have been able to observe deviations from this thermal pattern by seeing changes in the control routines captured by highly sensitive infrared cameras. The deviations came either in the form of minimal temperature fluctuations or temporal changes measured in milliseconds.
The setup with infrared cameras is initially being used to demonstrate that thermal monitoring is a fundamentally feasible step. In the future, sensors on the chip will take the place of the cameras. Such sensors already exist. But they are used only to prevent overheating. When teamed with neural networks, they will be able to identify thermal “anomalies” on their own and monitor the chip in real time.
Initially, they will be used primarily in industrial settings. Deviations are easier to spot here because generally static control routines are performed in this area. But the story is different for a smartphone. The sabotage threat and the amount of possible damage may be much less in this regard. But it can be assumed that it will become increasingly difficult to identify cyberattacks on industrial computers. After all, hackers who are “generally up to speed” can use smaller and slower malware to hide the thermal profile. But neural networks are capable of recognizing modified threats.