On-chip thermal sensors, neural networks protect against malware

Technology News | January 16, 2018

By Christoph Hammerschmidt

Over the past decade, the structure sizes in semiconductors has shrunk so strongly that the physical flow of a few electrons is now sufficient to execute software. This progress also has a downside: processors with structures of less than 10 nanometers are so sensitive that a targeted overload with incorrect control commands could trigger an artificial aging process that destroys them within a few days. A research group at the KIT is now working on an intelligent self-monitoring system in order to be able to ward off such attacks on industrial plants in the future.

The approach is based on the identification of thermal patterns in the normal operation of processors: “Each chip generates a specific thermal fingerprint,” explains Professor Jörg Henkel, who leads the research group at the Chair for Embedded Systems (CES) of the research institute based in Karlsruhe (Germany): “Calculations are carried out, data is stored in the working memory or retrieved from the hard disk. All these operations lead to brief heating and cooling in different areas of the processor.”

His research group observed this pattern with highly sensitive infrared cameras and was able to track changes in the control routine based on minimal temperature fluctuations or temporal deviations in the range of milliseconds. The experimental setup with infrared cameras served to prove the feasibility of such a thermal monitoring.

In the future, sensors on the chip will take over the function of the cameras. “There are already temperature sensors on the chips. They serve as overheating protection,” says Henkel. “We will increase the number of sensors and use them for cyber-security purposes for the first time. In addition, the scientists want to equip chips with neural networks that can identify thermal deviations independently and thus monitor the chip in real time.

Initially, it is likely that their technology will be used in practice in industrial plants, the researchers believe. In such environments, static control routines are typically executed, where deviations are easier to identify than in a smartphone. However, industrial computers are also concerned with a dynamic threat situation: “If the hackers know that the temperature is being monitored, they will adapt,” explains computer scientist Hussam Amrouch, who works with Jörg Henkel’s team: “They will write smaller or slower programs whose heating profiles are more difficult to detect. The neural networks should therefore be trained from the outset in such a way that they also recognize a modified threat.