Publications

National Sinmungo, a representative online civil complaint system in Korea, frequently receives complaints in which petitioners emotionally write personal attacks or profanity. This imposes excessive emotional labor and workload on public officials and is a major factor hindering efficient complaint handling. In this study, we propose Minwon 99, a large language model (LLM)-based system for profanity filtering and similar-case recommendation. Minwon 99 displays the toxicity level of each complaint and provides similar previously handled cases, thereby reducing public officials’ emotional labor and improving work efficiency. To evaluate its performance, we applied the filtering to 15,000 samples from the UnSmile dataset and measured its effectiveness. We also conducted a System Usability Scale (SUS) assessment and the Product Reaction Cards method with 22 incumbent public officials to examine the usability of Minwon 99 and its capacity to alleviate emotional strain.

Since the advent of large language models, prompt engineering now enables the rapid, low effort creation of diverse autonomous agents that are already in widespread use. Yet this convenience raises urgent concerns about the safety, robustness, and behavioural consistency of the underlying prompts—along with the pressing challenge of preventing those prompts from being exposed to user’s attempts. In this paper, we propose the “Doppelgänger method” to demonstrate the risk of an agent being hijacked, thereby exposing system instructions and internal information. Next, we define the "Prompt Alignment Collapse under Adversarial Transfer(PACAT)" level to evaluate the vulnerability to this adversarial transfer attack. We also propose a Caution for Adversarial Transfer(CAT) prompt to counter Doppelgänger method. The experimental results demonstrate that the Doppelgänger method can compromise the agent’s consistency and expose its internal information. In contrast, CAT prompts enable effective defense against this adversarial attack.

External human–machine interfaces (eHMIs) may be useful for communicating the intention of an autonomous vehicle (AV) to road users, but it is questionable whether an eHMI is effective in guiding the actual behavior of road users, as intended by the eHMI. To address this question, we developed a Unity game in which the player drove a conventional car and the AVs were operating with eHMIs. We examined the effects of different eHMI designs—namely, textual, graphical, and anthropomorphic—on the driving behavior of a player in a gaming environment, and compared it to one with no eHMI. Participants (N = 18) had to follow a specified route, using the typical keys for PC games. They encountered AVs with an eHMI placed on the rear window. Five scenarios were simulated for the specified routes: school safety zone; traffic island; yellow traffic light; waiting for passengers; and an approaching e-scooter. All scenarios were repeated three times (a total of 15 sessions per participant), and the eHMI was randomly generated among the four options. The behavior was determined by observing the number of violations in combination with keystrokes, fixations, and saccades. Their subjective evaluations of the helpfulness of the eHMI and their feelings about future AVs revealed their attitudes. Results showed that a total of 45 violations occurred, the most frequent one being exceeding the speed limit in the school safety zones (37.8%) when the eHMI was textual, anthropomorphic, graphical, and when there was no eHMI, in decreasing order; the next was collisions (33.3%), when the eHMI was anthropomorphic, none, or graphical. The rest were ignoring the red light (13.3%), crossing the stop line (13.3%), and violation of the central line (2.2%). More violations occurred when the eHMI was set to anthropomorphic, followed by no eHMI, graphical, and textual eHMI. The helpfulness of the five scenarios scored high (5.611 to 6.389) on a seven-point Likert scale, and there was no significant difference for the scenarios. Participants felt more positive about the future of AVs after their gaming experience (p = 0.049). We conclude that gazing at unfamiliar and ambiguous information on eHMIs may cause a loss of driver attention and control. We propose an adaptive approach in terms of timing and distance depending on the behavior of other road users.

In disasters and emergency situations, it is necessary to immediately prescribe fluids such as 0.9% normal saline and glucose, as soon as a patient is rescued. However, the current gravity-based fluid infusion method requires a certain height, which reduces rescue efficiency and causes secondary medical accidents that injure rescuers or patients. In this study, we designed a non-powered medicine injector that uses elastic force and negative air pressure, rather than gravity, to address these issues, taking into account the challenges of using electricity in disaster situations. To design a non-powered medicine injector, we analyzed the advantages and disadvantages of existing products. Then we conducted in-depth interviews with four experts working in the field of emergency medicine to identify the problem, and based on this, we first designed a prototype of the product. To verify the feasibility of the prototype design, we conducted a few experiments comparing existing methods. And finally we proposed a final design of non-powered medicine injector.

Can autonomous vehicles (AVs) interact emotionally with other road users to influence their behavior? To address this question, we developed an AV with external human–machine interfaces (eHMIs) in Roblox. The anthropomorphic facial expressions of the AV which conveyed the emotional state of the AV were designed. We assumed that people would behave better with an anthropomorphic AV, resulting in better traffic flow. In this paper, the late-breaking results of our pilot A/B test conducted to select appropriate emojis are presented. The results show that participants felt positive about the anthropomorphic eHMI because of the intuitiveness of the emojis, which made them feel better and less stressed in traffic situations.