Research Bits: Nov. 7

ADC side-channel attacks

Researchers at MIT propose two ways to protect analog-to-digital converters (ADCs) from power and electromagnetic side-channel attacks.

The researchers first investigated the side-channel attacks that could be used against ADCs. Power attacks usually involve an attacker soldering a resistor onto the device’s circuit board to measure its power usage. An electromagnetic attack is noninvasive, using an electromagnetic probe that can monitor electric current without touching the device. They found that electromagnetic attacks were just as effective as power attacks, even when the probe was held 1 centimeter away from the chip. They note this suggests a hacker could use this attack to steal private data from an implantable medical device.

“Side-channel attacks are always a cat and mouse game. If we hadn’t done the work, the hackers most likely would have come up with these methods and used them to attack analog-to-digital converters, so we are preempting the action of the hackers,” said Hae-Seung Lee, the advanced television and signal processing professor of electrical engineering and director of the Microsystems Technology Laboratories at MIT.

To address this, the teams added randomization to the ADC conversion process.

“A common type of ADC sets a threshold in the center of its voltage range and uses a circuit called a comparator to compare the input voltage to the threshold. If the comparator decides the input is larger, the ADC sets a new threshold in the top half of the range and runs the comparator again. This process continues until the unknown range becomes so small it can assign a digital value to the input,” explained Adam Zewe of the MIT News Office. “The ADC typically sets thresholds using capacitors, which draw different amounts of electric current when they switch. An attacker can monitor the power supplies and use them to train a machine-learning model that reconstructs output data with surprising accuracy.”

In the first countermeasure, which is suited for low-power applications like smart devices, a random number generator decides when each capacitor switches, making it harder for an attacker to correlate power supplies with output data. This technique also keeps the comparator running constantly to prevent an attacker from determining when each stage of the conversion began and ended. The method required 14% more chip area.

“The idea is to split up what would normally be a binary search process into smaller chunks where it becomes difficult to know what stage in the binary search process you are on. By introducing some randomness into the conversion, the leakage is independent from what the individual operations are,” said Maitreyi Ashok, a graduate student at MIT.

In the second method, which is more suited for high-speed applications like video processing, the ADC randomizes the starting point of the conversion process. This method uses two comparators and an algorithm to randomly set two thresholds instead of one, so there are millions of possible ways an ADC could arrive at a digital output. The researchers said this makes it nearly impossible for an attacker to correlate a power supply waveform to a digital output. Additionally, it did not require more chip area and enables it to run almost as fast as a standard ADC.

“For the past half-century of ADC research, people have focused on improving the power, performance, or area of the circuit. We’ve shown that it is also extremely important to consider the security side of ADCs. We have new dimensions for designers to consider,” said Ruicong Chen, a graduate student at MIT.

Next, the researchers plan to use the methods to develop detection-driven chips, where protection only turns on when the chip detects a side-channel attack, to boost energy efficiency while maintaining security.

AI image sensor

Researchers at University of Central Florida, Stony Brook University, Korea Basic Science Institute, and the Air Force Research Lab developed an image sensor that can mimic the retina and recognize what it sees. It is capable of sensing ultraviolet, visible, and infrared light.

“Today, everything is discrete components and running on conventional hardware. And here, we have the capacity to do in-sensor computing using a single device on one small platform,” said Tania Roy, an assistant professor in UCF’s Department of Materials Science and Engineering and NanoScience Technology Center. “We had devices, which behaved like the synapses of the human brain, but still, we were not feeding them the image directly. Now, by adding image sensing ability to them, we have synapse-like devices that act like ‘smart pixels’ in a camera by sensing, processing and recognizing images simultaneously.”

The device uses nanoscale surfaces made of molybdenum disulfide and platinum ditelluride to allow for multi-wavelength sensing and memory.

The technology is very small, with hundreds of the devices fitting on a one-inch-wide chip. Credit: University of Central Florida

Molla Manjurul Islam, a doctoral student in UCF’s Department of Physics, highlights the importance of multi-wavelength capability. “If you are in your autonomous vehicle at night and the imaging system of the car operates only at a particular wavelength, say the visible wavelength, it will not see what is in front of it. But in our case, with our device, it can actually see in the entire condition. There is no reported device like this, which can operate simultaneously in ultraviolet range and visible wavelength as well as infrared wavelength, so this is the most unique selling point for this device.”

The researchers tested the device’s accuracy by having it sense and recognize a mixed wavelength image — an ultraviolet number “3” and an infrared part that is the mirror image of the digit that were placed together to form an “8.” They demonstrated that the technology could discern the patterns and identify it both as a “3” in ultraviolet and an “8” in infrared.

“We got 70 to 80% accuracy, which means they have very good chances that they can be realized in hardware,” said Adithi Krishnaprasad, a doctoral student in UCF’s Department of Electrical and Computer Engineering.

The researchers expect the technology could be available in five to ten years.

5G transceiver

Researchers from Tokyo Institute of Technology propose a phased-array beamformer for mmWave 5G base stations based upon the Doherty amplifier and digital predistortion (DPD).

The team modified the conventional Doherty amplifier design, which offers good power efficiency and suitability for signals with a high peak-to-average ratio, and produced a bi-directional amplifier that allows the same circuit to both amplify a signal to be transmitted and a received signal with low noise.

“Our proposed bidirectional implementation for the amplifier is very area-efficient. Additionally, thanks to its co-design with a wafer-level chip-scale packaging technology, it enables low insertion loss. This means that less power is lost as the signal traverses the amplifier,” said Kenichi Okada, a professor at Tokyo Tech.

However, the Doherty amplifier can exacerbate nonlinearity problems that arise from mismatches in the elements of the phased-array antenna.

To address this, the team employed the DPD technique, which involves distorting the signal before transmission to effectively cancel out the distortion introduced by the amplifier. Their implementation, unlike conventional DPD approaches, used a shared look-up table (LUT) for all antennas, minimizing the complexity of the circuit.

Second, they introduced inter-element mismatch compensation capabilities to the phased array, improving its overall linearity. “We compared the proposed device with other state-of-the-art 5G phased-array transceivers and found that, by compensating the inter-element mismatches in the shared-LUT DPD module, ours demonstrate a lower adjacent channel leakage and transmission error,” said Okada. “Hopefully, the device and techniques described in this study will let us all reap the benefits of 5G NR sooner!”