Jihun Kim: Pioneering Resilient Control Systems
- Miguel Virgen, PhD Student in Business
- 15 hours ago
- 4 min read
PhD student Jihun Kim of UC Berkeley’s Industrial Engineering & Operations Research department, alongside his advisor Professor Javad Lavaei, has earned recognition as a Best Student Paper finalist at the 2025 American Control Conference (ACC). Their paper, “Prevailing Against Adversarial Noncentral Disturbances: Exact Recovery of Linear Systems with the $l_1$-Norm Estimator,” was selected as one of the top six submissions out of roughly 1,000 accepted each year. ACC is a premier annual forum for control systems engineering that convenes a global community of researchers and practitioners (a2c2.org). In this work, Kim and Lavaei tackle a fundamental challenge: how to accurately learn a linear system’s dynamics in the presence of adversarial disturbances (ieor.berkeley.edu). Such resilient learning techniques could improve the reliability of modern engineered systems, including autonomous vehicles and smart power grids (ieor.berkeley.edu). For context, ACC 2025 is expected to host roughly 1,300 participants (a2c2.org), underscoring the broad visibility of Kim’s research.
Academic Journey: From Seoul to Berkeley
Kim’s academic journey laid a strong foundation for his research achievements. He earned Bachelor of Science degrees in Industrial Engineering and Statistics from Seoul National University (SNU) in 2022, graduating with a 4.22/4.3 GPA and ranking first in his class (lavaei.ieor.berkeley.edu). He then began his doctoral studies at UC Berkeley in 2022, becoming a fourth-year PhD student in the IEOR department under the guidance of Professor Javad Lavaei (lavaei.ieor.berkeley.edusites.google.com). On his personal profile, Kim describes his research interests as the theoretical understanding of dynamic systems and learning-based controls under unforeseen attacks, along with optimization techniques in online learning and control (sites.google.com). This blend of theoretical control, statistics, and optimization expertise prepared Kim to address the complexities of resilient system identification. In addition, Kim gained industry experience through internships in sports analytics and semiconductor image analysis during his undergraduate years (sites.google.com), demonstrating a strong applied data science background.
ACC Finalist Research: Learning Under Adversarial Disturbances
In their ACC finalist paper, Kim and Lavaei analyze system identification under challenging conditions. They consider linear dynamical systems where the process disturbances are not only sub-Gaussian and correlated, but potentially adversarially chosen. In the first scenario, disturbances have a nonzero mean (so-called “noncentral” disturbances), which makes the ordinary least-squares (OLS) estimator fail. They prove that an $l_1$-norm estimator – which minimizes the sum of absolute errors – can still correctly identify the true system, provided that each disturbance sample is equally likely to be positive or negative. Next, they study the case of sporadic large attacks: assuming that attacks occur with probability less than 0.5 at any given time, they show that the $l_1$-norm estimator “prevails against any adversarial noncentral disturbances” and achieves exact recovery of the system parameters in finite time. In essence, even infrequent malicious disturbances cannot derail learning the correct linear model. These findings integrate tools from robust statistics and adversarial machine learning into control theory, ensuring that learned models remain valid even under malicious data corruption.
Technical Insight: Why the $l_1$-Norm Works
The power of the $l_1$-norm estimator lies in its robustness to outliers and bias. Unlike least-squares (which minimizes the sum of squared errors and can be skewed by large disturbances), the $l_1$ approach minimizes the sum of absolute residuals. This means that extremely large disturbances have a limited impact on the estimated model. When noise samples have no net bias – that is, positive and negative errors occur equally – the $l_1$ minimization effectively centers on the true system parameters. Even under attack, if more than half of the measurements are uncompromised, the $l_1$ estimator essentially “ignores” the outliers (much like taking a median) and converges to the correct parameters. In technical terms, solving the $l_1$-minimization finds parameters that fit the median behavior of the data, naturally discarding extreme outliers. This insight explains why Kim and Lavaei’s method can recover the exact system despite strong, infrequent disturbances.
Implications for Modern Engineering Systems
Kim and Lavaei’s findings promise to bolster the resilience of safety-critical control systems. As the UC Berkeley announcement notes, this work has “significant implications for improving the resilience and reliability of modern engineering systems, from autonomous vehicles to smart grids". For example, an autonomous vehicle learns its own dynamics from sensor data; if a malicious agent or sensor glitch injects false readings occasionally, a resilient estimator like the one developed by Kim will still identify the true vehicle model. In power networks, system parameters must be estimated online even under large demand spikes or cyberattacks; the $l_1$-norm method ensures that sporadic large errors do not mislead the identification. Similar ideas extend to other applications: networked IoT sensors can use robust identification to resist data tampering, and robotic controllers can remain accurate even if some sensor inputs are corrupted. By enabling exact recovery of the linear model under adversarial conditions, this research paves the way for more secure and robust control in real-world systems.
LinkedIn Profile:
Keywords:
Jihun Kim control systems research, adversarial disturbances system identification, l1-norm estimator linear systems, ACC 2025 control finalist, resilient engineering systems.
