Technology
Exploring cybersecurity needs for energy-efficient analog computing
Baskin Engineering hardware security expert Dustin Richmond will study vulnerabilities in compute-in-memory architectures.
Getty Images/ JERO SenneGs
Press Contact
Digital computers use complex algorithms of 1s and 0s to run much of the computing integrated into our day-to-day lives, from using a calculator to asking questions of AI. They’ve been refined by decades of engineering to perform extraordinarily well, but can have high energy costs, especially for modern-day AI systems.
Analog computing systems can have major energy savings, so some are proposing this to improve effeciency—but the cybersecurity risks of this approach for modern computing remain largely unknown. University of California, Santa Cruz, Assistant Professor of Computer Science and Engineering Dustin Richmond will study and design mitigations against potential attacks on analog compute-in-memory systems with the support of a $743,000 National Science Foundation CAREER award.
“We’ve operated in the digital domain for the last 40 or 50 years, but the analog domain is an opportunity for us to be more efficient,” Richmond said. “We just don’t yet understand the security implications of that.”
Computing in memory
Analog compute systems perform calculations directly via physical laws that govern the movement of charged particles through a chip. This is a more efficient process than digital systems, in which algorithms have to work through many step-by-step logic operations to get the same result that physics does in one smooth motion.
While some engineers have designed systems that perform compute near-memory, there are no widespread commercial compute-in-memory architectures deployed today, in which data is processed directly by the element where it’s stored. But Richmond, who studies computer architecture and hardware design at the Baskin School of Engineering, anticipates that this kind of systems will become relevant in the near future because of the opportunities for efficiency as it becomes harder to pack more digital compute onto existing chips. This approach could be used for a wide range of computing needs like signal processing and AI.
Mitigating vulnerabilities
Although they have energy efficiency benefits, analog compute systems are vulnerable to a problem called a charge migration. This is a phenomenon similar to how, in old-school TVs or computers, words or images could be permanently burned onto a screen after an image is displayed for a long period of time. In essence, the physical pixels of the screen become overused and degraded. In analog compute systems, overuse of a certain computing process can quickly physically degrade the chip in ways that indicate details of the process itself.
“In this case, it could be that neural network weights are used for so long they affect the actual current that’s being used,” Richmond said. “Our research has shown this can happen much more quickly than people expect, possibly in hours.”
While chips can recover from these channel migration effects, it’s not clear how long this recovery period takes. Furthermore, the recovery effect could also leak information. In the meantime, bad actors could take advantage of these effects to extract vulnerable information.
Richmond’s project will provide a deep study of channel migration attacks and propose methods for mitigating them, likely in the form of chip design recommendations. He will examine three major compute-in-memory approaches, SRAM, eDRAM, and ReRAM, to gain key insights.
“When we know more about the timescales, and about the effects and outcomes, we can suggest mitigations against these attacks—but we have to study them in the first place,” Richmond said.
A major component of the NSF’s CAREER awards is to create educational opportunities for students. Richmond aims to expand access to introductory computer engineering education for students at schools that lack existing resources by creating new programs or developing open source educational tools.