That being said, researchers at Ruhr University Bochum, Germany, and the Max Planck Institute for Security and Privacy (MPI-SP) in Bochum want to do away those attacks as they are exploring methods of detecting these hardware Trojans. To get there the researchers compared construction plans for chips with electron microscope images of real chips and had an algorithm search for differences. They detected deviations in 37 out of 40 cases.
The researchers released all images of the chips, the design data as well as the analysis algorithms online so other research groups can use the data to conduct further studies.
These days, electronic chips are integrated into countless objects. They are more often than not designed by companies that don’t operate their own production facilities. The construction plans are therefore sent to highly specialized chip factories for production.
“It’s conceivable that tiny changes might be inserted into the designs in the factories shortly before production that could override the security of the chips,” Becker said. “In extreme cases, such hardware Trojans could allow an attacker to paralyze parts of the telecommunications infrastructure at the push of a button.”
Becker and Puschner’s team analyzed chips produced in the four modern technology sizes of 28, 40, 65 and 90 nanometers. For this purpose, they collaborated with Moos, who had designed several chips as part of his PhD research at Ruhr University Bochum and had them manufactured. Thus, the researchers had both the design files and the manufactured chips at their disposal.
They obviously couldn’t modify the chips after the fact and build in hardware Trojans. And so they employed a trick: Rather than manipulating the chips, Moos changed his designs retroactively in order to create minimal deviations between the construction plans and the chips. Then, the Bochum researchers tested if they could detect these changes without knowing what exactly they had to look for and where.
In the first step, the team at Ruhr University Bochum and MPI-SP had to prepare the chips using complex chemical and mechanical methods in order to take several thousand images of the lowest chip layers with a scanning electron microscope. These layers contain several hundred thousand of standard cells that carry out logical operations.

w/art: Researchers took thousands of microscopic images of microchips. Pictured here is such a chip in a golden chip package. The chip area that was inspected only measures about two square millimeters.Source: RUB, Marquard
“Comparing the chip images and the construction plans turned out to be quite a challenge, because we first had to precisely superimpose the data,” Puschner said. In addition, every little impurity on the chip could block the view of certain sections of the image. “On the smallest chip, which is 28 nanometers in size, a single speck of dust or a hair can obscure a whole row of standard cells,” he said.
The researchers used image processing methods to carefully match standard cell for standard cell and looked for deviations between the construction plans and the microscopic images of the chips.
“The results give cause for cautious optimism,” Puschner said.
For chip sizes of 90, 65 and 40 nanometers, the team successfully identified all modifications. The number of false-positive results totaled 500, i.e. standard cells were flagged as having been modified, although they were in fact untouched.
“With more than 1.5 million standard cells examined, this is a very good rate,” Puschner said.
It was only with the smallest chip of 28 nanometers the researchers failed to detect three subtle changes.
A better recording quality could remedy this problem in the future.
“Scanning electron microscopes do exist that are specifically designed to take chip images,” Becker said.
Moreover, using them in a clean room where contamination can be prevented would increase the detection rate even further.
“We also hope that other groups will use our data for follow-up studies,” Becker said. “Machine learning could probably improve the detection algorithm to such an extent that it would also detect the changes on the smallest chips that we missed.”

