UK IC Postdoctoral Research Fellowships
2022-2024
Machine learning is slow and consumes massive amounts of energy. This is caused by the repeated transfer of large quantities of data in traditional computer architectures, where memory and computing are separated. This problem is only getting worse because state-of-the-art machine learning models, like GPT-4, keep increasing in size.
Computer hardware based on physical computing might address this issue. As opposed to digital computers, circuits made up of analogue devices (like variable resistors) may be able to (1) store data more densely and (2) perform computations using physics without even moving the data. This could improve energy efficiency and reduce computation time by orders of magnitude. Unfortunately, analogue circuitry is less precise and less predictable compared to digital devices, such as transistors, thus many challenges still remain.
One of the less understood areas is the cybersecurity of analogue circuits that are used as hardware accelerators for machine learning. Even in traditional machine learning, there is a risk of attacks—for example, if an autonomous car uses smart classifiers to distinguish between different traffic signs, an adversary may be able to perturb the inputs to the camera, resulting in a misclassification and potentially, an accident. The capabilities of such attacks are even less understood when analogue hardware is used—can nonideal behaviours of these devices be exploited by an adversary, or could they actually make it more difficult to construct an effective attack because the system is less predictable?
Dr Joksas aims to answer these questions through a combination of theoretical and simulation-based work. Using experimental data from analogue devices called memristors, he is investigating the feasibility of adversarial attacks on machine learning models running on analogue hardware accelerators. Additionally, he is designing these systems, which may complement traditional compute hardware in the future, to be more robust using circuit design and algorithmic techniques.
Personal website: https://yoshke.org
Linkedin: https://www.linkedin.com/in/joksas
Twitter: https://twitter.com/DovydasJoksas
![](/media/2fzf1mtf/dovydas-joksas.jpg)
Related content
View all programmesSupport for Research
The Academy runs a number of grants to support excellent researchers carry out engineering activities and to enable clo…
UK IC Postdoctoral Research Fellowships
These Fellowships are designed to promote unclassified basic research in areas of interest to the intelligence, securit…