-
- [25.04.30] Data Intelligence and Learning Lab (Professor: Jongwook Lee) Published 3 SIGIR 2025 Papers
- The Data Intelligence and Learning (DIAL, Professor: Jong-Wook Lee) lab has had three papers accepted for publication at SIGIR 2025, the world's most prestigious information retrieval conference, and will present them in Padua, Italy in July. [List of Papers] 1. Why is Normalization Necessary for Linear Recommenders? (SIGIR'25) 2. Linear Item-Item Models with Neural Knowledge for Session-based Recommendation (SIGIR'25) 3. DIFF: Dual Side-Information Filtering and Fusion for Sequential Recommendation (SIGIR'25) Study 1: Seongmin Park, Mincheol Yoon, Hye-young Kim, Jongwuk Lee, “Why is Normalization Necessary for Linear Recommenders?”, The 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2025 This study focuses on the fact that linear autoencoder (LAE)-based recommendation systems have comparable recommendation performance and fast inference speed to neural network-based models despite their simple structure. However, LAE faces two structural limitations: popularity bias, which over-recommends popular items, and neighborhood bias, which relies excessively on local correlations between items. To address these issues, in this paper, we propose a novel normalization method, Data-Adaptive Normalization (DAN), that can be applied to the LAE model. DAN is a normalization technique designed to flexibly control two biases depending on the characteristics of the data, and consists of two core components: (i) item-adaptive normalization and (ii) user-adaptive normalization. First, item-adaptive normalization controls the influence of popular items through the regularization parameter α and provides a denoising effect to LAE. This allows LAE to significantly improve the recommendation performance for unpopular items (tail items) by moving away from the performance centered on popular items (i.e., popularity bias) that the existing LAE mainly recommends. Second, user-adaptive normalization controls the neighborhood bias using the parameter β. This technique suppresses high-frequency components and preserves important low-frequency components, thereby helping to better reflect the overall global pattern rather than local correlations. The effectiveness of DAN is experimentally verified on six representative recommendation datasets (ML-20M, Netflix, MSD, Gowalla, Yelp2018, Amazon-book). LAE models (LAE_DAN, EASE_DAN, RLAE_DAN) applied with DAN showed consistent performance improvement over the existing LAE model on all datasets, and recorded performance improvements of up to 128.57% and 12.36% in tail items and unbiased evaluation, respectively. DAN also showed superior performance compared to the state-of-the-art collaborative filtering models. In addition, Figure 1 (Case study) shows the recommendation results for a specific user according to the regularization method, and the following observations were made: (1) LAE without regularization (W/O) recommends only five highly popular action movies even though the user watched three romantic movies. On the other hand, the three regularization methods (RW, Sym, DAN) effectively reflect user preferences by recommending “Step Up 2” as the top item related to “Step Up 1” viewed by the user. (2) DAN provides the most balanced recommendation that maintains user preferences while appropriately mitigating popularity bias. RW regularization still has strong popularity bias with 4 out of 5 items being popular. Sym regularization overly mitigates popularity bias with 4 out of 5 items being unpopular. DAN recommends the items most relevant to user preferences while balancing popular and unpopular items. Figure 1: Interaction history of user #91935 on the ML-20M dataset and the top-5 recommendation list of the four regularization methods. The red border is the head (top 20%) item, and the blue border is the tail (bottom 80%) item. Furthermore, this study analyzes how the effect of the regularization parameters (α, β) varies depending on the Gini index and homophily characteristics of the dataset, and also presents guidelines for setting parameters suitable for each dataset. Through this, it is shown that the proposed DAN technique can be established as a general and practical solution that can precisely control bias depending on the data characteristics. For more information about this paper, please refer to the following address. https://dial.skku.edu/blog/2025_dan Study 2: Minjin Choi, Sunkyung Lee, Seongmin Park, Jongwuk Lee, “Linear Item-Item Models with Neural Knowledge for Session-based Recommendation”, The 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2025 This study focuses on the problem of session-based recommendation (SBR), which predicts the next action based on the user’s current session interaction. The SBR field is largely divided into two paradigms. One is a neural network-based model that is strong in capturing complex sequential transition patterns, and the other is a linear item-item model that effectively learns co-occurrence patterns between items and provides fast inference speed. However, each paradigm is specialized in capturing different types of item relationships, and an effective integrated method to simultaneously achieve high accuracy of neural network models and efficiency of linear models is still lacking. Against this backdrop, in this paper, we propose a novel SBR model, LINK (Linear Item-Item model with Neural Knowledge), which effectively integrates knowledge from linear models and neural network models. LINK aims to achieve both high accuracy and fast inference speed by combining two types of knowledge within a single unified linear framework. To this end, LINK includes two core components. (i) LIS (Linear knowledge-enhanced Item-item Similarity model) enhances the linear model’s ability to capture item similarity (co-occurrence) and learns high-dimensional relationships between sessions through self-distillation technique. (ii) NIT (Neural knowledge-enhanced Item-item Transition model) effectively injects neural network knowledge into linear models through a unique method that distills complex sequential transfer knowledge from pre-trained arbitrary neural network models and utilizes it as a regularization term when learning linear models. As shown in Figure 2, the effectiveness of the LINK model has been verified through extensive experiments using six real-world SBR datasets, including Diginetica, Retailrocket, and Yoochoose. The experimental results show that LINK consistently and significantly improves performance (up to 14.78% in Recall@20 and up to 11.04% in MRR@20) compared to existing state-of-the-art linear SBR models (such as SLIST and SWalk) on all datasets. This demonstrates that LINK successfully overcomes the limitations of linear models by incorporating neural network knowledge. In addition, LINK maintains the high inference efficiency (up to 813 times fewer FLOPs), which is a key advantage of linear models, while showing competitive or superior prediction accuracy compared to complex state-of-the-art neural network models. Further analysis shows that linear models are strong in the relationship between unpopular items, while neural network models are strong in the complex patterns of popular items, and LINK effectively combines these two strengths to achieve balanced performance. Figure 2: Comparison of Accuracy (Recall@20) and Inference Operations (FLOPs) In conclusion, LINK presents a novel hybrid approach that provides a practical trade-off between accuracy and efficiency in the SBR field. In particular, the NIT component provides the flexibility to leverage knowledge from various models without being bound to a specific neural network architecture, making it a practical solution that can continuously improve performance as neural network models evolve in the future. For more information about this paper, please refer to the following address: https://dial.skku.edu/blog/2025_link Study 3: Hye-young Kim, Minjin Choi, Sunkyung Lee, Ilwoong Baek, Jongwuk Lee, “DIFF: Dual Side-Information Filtering and Fusion for Sequential Recommendation”, The 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2025 This study proposes a side-information integrated sequential recommendation (SISR) model that utilizes additional information such as category and brand in sequential recommendation, which predicts the next preferred item based on the user’s past consumption history. The proposed model, Dual Side-Information Filtering and Fusion (DIFF), removes noise in the user sequence and effectively fuses various attribute information to achieve more precise and expressive user preference modeling. DIFF includes the following three core techniques: Figure 3: Frequency signals and fusion techniques of sequential recommendation systems that integrate additional information (1) Frequency-based noise filtering: DIFF performs frequency domain transformation to remove signals that are not related to actual user preferences, such as accidental clicks or short-term interests. After converting the item ID and each attribute sequence to the frequency domain, it removes irregular or low-importance frequency components. This allows us to strengthen only the core signals that reflect actual user preferences, and enables more sophisticated noise removal by applying filtering to multiple sequences. (2) Dual multi-sequence fusion: DIFF utilizes intermediate fusion and early fusion methods, which have different strengths, together to effectively integrate the denoised sequences. We note that previous studies have tended to limit or exclude the use of early fusion methods due to concerns about information invasion, which has led to the overlooking of the ability to model correlations between various attributes. DIFF learns sophisticated user representations that encompass both IDs and attributes by integrating multidimensional attribute information through early fusion and supplementing ID-centric preference learning through intermediate fusion. Through the complementary combination of the two fusion methods, DIFF can effectively capture not only the overall structure of user tastes but also detailed attribute preferences. (3) Representation alignment to prevent information invasion: Item IDs and each attribute embedding are located in different representation spaces. Therefore, early fusion that combines them with a simple fusion function (e.g. summation, concatenation, gating) may cause information invasion, where specific information is overly emphasized or distorted. To prevent this, DIFF designs an alignment loss to make the vector spaces of item IDs and attribute embeddings close together, thereby maintaining appropriate differences while sharing meaning. DIFF has been validated on four representative public benchmark datasets (Yelp, Beauty, Toys, Sports), and has achieved superior performance in all indicators compared to existing state-of-the-art sequential recommendation models. In particular, it has demonstrated a new state-of-the-art performance by recording performance improvements of up to 14.1% and 12.5% on Recall@20 and NDCG@20, respectively. In addition, DIFF's robustness against noise is very remarkable. Considering noise in a realistic usage environment such as accidental clicks and temporary changes in interest in the user sequence, we conducted noise simulation experiments by randomly replacing items in the test sequence. As a result, DIFF showed the least performance degradation compared to other models even under low noise conditions of 5%, and maintained stable high performance even under high noise conditions of 25%. For more information about this paper, please refer to the following address: https://dial.skku.edu/blog/2025_diff
-
- 작성일 2025-04-30
- 조회수 82
-
- [25.04.30] Professor Lee Ji-hyung elected as new president of the Korean Society for Artificial Intelligence
- Professor Lee Ji-hyung Elected as New President of the Korean Society for Artificial Intelligence Professor Lee Ji-hyung of the Department of Software Elected as the 5th President of the Korean Society for Artificial Intelligence on April 16. Professor Lee has been serving as a director of the society and has contributed to the academic and technological development of the artificial intelligence field. In 2022, he served as the organizing committee chairman of the ‘Korean Society for Artificial Intelligence & Naver Fall Joint Academic Conference’ and played a major role in revitalizing academic exchange. Professor Lee Ji-hyung received the Sungkyunkwan Family Award in the Educational Achievement category from our school in 2019, and in 2022, he received the Byeon Jeong-nam Academic Award from the Korean Society for Intelligent Systems in recognition of his contributions to the development of artificial intelligence. In 2023, he received the Minister of Science and ICT Award in recognition of his contributions to fostering artificial intelligence talent, and continues to make outstanding achievements in all areas of education and research. Meanwhile, the Korean Artificial Intelligence Society, established in 2016, is actively promoting academic exchanges and industry-academia-research cooperation activities for the purpose of research and education in artificial intelligence and related fields such as computer vision, pattern recognition, natural language processing, bioinformatics, brain cognitive computing, and machine learning. Through this, it is playing a pivotal role in the development and spread of domestic artificial intelligence technology.
-
- 작성일 2025-04-30
- 조회수 98
-
- [25.03.28] Security Engineering Laboratory (SecLab) under Professor Kim Hyung-sik – Paper Accepted for Publication at...
- Security Engineering Laboratory (SecLab) at SKKU (Advisor: Kim Hyung-sik, https://seclab.skku.edu) – "Open Sesame! On the Security and Memorability of Verbal Passwords" Accepted for IEEE Symposium on Security and Privacy (S&P) 2025 The paper "Open Sesame! On the Security and Memorability of Verbal Passwords," conducted by Ph.D. candidate Kim Eun-soo and Professor Kim Hyung-sik at the Security Engineering Laboratory, has been accepted for publication at the IEEE Symposium on Security and Privacy (S&P) 2025, one of the most prestigious conferences in the field of computer security. The study was conducted in collaboration with Professor Kim Doo-won of the University of Tennessee and alumnus Lee Ki-ho from the Security Engineering Laboratory (currently at ETRI). The research quantitatively analyzed the security and memorability of verbal passwords through two large-scale user experiments, demonstrating that verbal passwords offer a practical and secure alternative to traditional text-based passwords by overcoming their inherent limitations. In the first user experiment, verbal passwords freely generated by 2,085 participants were evaluated for both short-term and long-term memorability as well as security. Security testing conducted using the PassphraseGPT model—trained on over 20 million common English phrases—revealed that approximately 39.76% of the user-generated verbal passwords could be predicted within one billion guess attempts. In the second experiment, involving 600 participants, a password creation policy that enforced a minimum word count and incorporated a blocklist was implemented. This approach significantly improved security while maintaining ease of memorability. In long-term memory tests, 65.6% of users in the verbal password group were able to successfully recall their passwords, compared to 54.11% for text-based passwords. Moreover, the proportion of verbal passwords susceptible to guessing attacks was lower than that of text passwords, indicating a stronger resistance to such attacks. This research has been highly acclaimed for demonstrating that verbal passwords provide a practical and secure alternative to text-based passwords in scenarios where keyboard input is either impossible or inconvenient—such as with smart assistants, wearable devices, in-vehicle systems, and VR/AR environments. The study will be presented in May 2025 in San Francisco, California, USA. Abstract Despite extensive research on text passwords, the security and memorability of verbal passwords—spoken rather than typed—remain underexplored. Verbal passwords hold significant potential for scenarios where keyboard input is impractical (e.g., smart speakers, wearables, vehicles) or users have motor impairments that make typing difficult. Through two large-scale user studies, we assessed the viability of verbal passwords. In our first study (N = 2,085), freely chosen verbal passwords were found to have a limited guessing space, with 39.76% cracked within 10^9 guesses. However, in our second study (n = 600), applying word count and blocklist policies for verbal password creation significantly enhanced verbal password performance, achieving better memorability and security than traditional text passwords. Specifically, 65.6% of verbal password users (under the password creation policy using minimum word counts and a blocklist) successfully recalled their passwords in long-term tests, compared to 54.11% for text passwords. Additionally, verbal passwords with enforced policies exhibited a lower crack rate (6.5%) than text passwords (10.3%). These findings highlight verbal passwords as a practical and secure alternative for contexts where text passwords are infeasible, offering strong memorability with robust resistance to guessing attacks.
-
- 작성일 2025-03-28
- 조회수 427
-
- [25.03.26] Security Engineering Laboratory (Advisor: Kim Hyung-sik) – Two Papers Accepted for Oral Sessions at The Web..
- The Security Engineering Laboratory, under the supervision of Professor Kim Hyung-sik, in collaboration with Professor Kim Doo-won from the University of Tennessee, has had two research papers accepted for oral sessions at The Web Conference (WWW) 2025, one of the premier international conferences in the web domain. In this research, alumnus Lee Ki-ho, a former member of the Security Engineering Laboratory (currently at ETRI), participated as a visiting researcher at the University of Tennessee and collaborated with Professor Kim Hyung-sik. Both papers, based on extensive empirical data, quantitatively analyze the characteristics and structures of phishing attacks. They have been highly acclaimed for providing a fundamental understanding of phishing attacks and proposing new countermeasures. The presentations are scheduled to take place in May 2025 in Sydney, Australia. Paper 1. 7 Days Later: Analyzing Phishing-Site Lifespan After Detected This paper presents an empirical study analyzing the lifetime and evolution of phishing sites after detection. Over a period of five months, 286,237 phishing URLs were tracked at 30-minute intervals to examine the attack patterns of phishing sites, shedding light on why the effectiveness of conventional phishing detection strategies is diminishing. Phishing sites have a short lifespan—with an average survival time of 54 hours and a median of 5.46 hours—highlighting the limitations of training and detection approaches. For instance, Google Safe Browsing detects phishing sites, on average, 4.5 days after their emergence; however, 84% of phishing sites cease operations before detection, demonstrating the inherent limitations of such detection methods. Paper 2. What's in Phishers: A Longitudinal Study of Security Configurations in Phishing Websites and Kits This paper presents a systematic analysis of phishing infrastructure by comprehensively examining the security configurations and structural vulnerabilities based on a combined dataset of 906,731 phishing websites and 13,344 phishing kits collected over a period of 2 years and 7 months. The study has attracted attention for proposing a proactive strategy that leverages the structural weaknesses of phishing sites to neutralize the attack infrastructure, thereby moving away from traditional passive detection and blocking methods and towards an early shutdown approach for phishing sites.
-
- 작성일 2025-03-28
- 조회수 463
-
- [25.03.17] IEEE S&P 2025 Paper Acceptance Announcement from Professor Lee Ho-jun’s Research Laboratory (SSLab)
- [IEEE S&P 2025 Acceptance Announcement – SSLab, Professor Hojoon Lee] The paper from the System Security Laboratory (SSLab), under the supervision of Professor Hojoon Lee, has been accepted for publication at IEEE S&P 2025, one of the four premier international conferences in the security field. The paper is scheduled for presentation in May in San Francisco, California, USA. Title: IncognitOS: A Practical Unikernel Design for Full-System Obfuscation in Confidential Virtual Machines Authors: Kha Dinh Duy, Jaeyoon Kim, Hajeong Lim, Hojoon Lee Summary: Recent works have repeatedly proven the practicality of side-channel attacks in undermining the confidentiality guarantees of Trusted Execution Environments such as Intel SGX. Meanwhile, the trusted execution in the cloud is witnessing a trend shift towards confidential virtual machines (CVMs). Unfortunately, several side-channel attacks have survived the shift and are feasible even for CVMs, along with the new attacks discovered on the CVM architectures. Previous works have explored defensive measures for securing userspace enclaves (i.e., Intel SGX) against side-channel attacks. However, the design space for a CVM-based obfuscation execution engine is largely unexplored. This paper proposes a unikernel design named IncognitOS to provide full-system obfuscation for CVM-based cloud workloads. IncognitOS fully embraces unikernel principles such as minimized TCB and direct hardware access to render full-system obfuscation feasible. IncognitOS retrofits two key OS components, the scheduler and memory management, to implement a novel adaptive obfuscation scheme. IncognitOS's scheduling is designed to be self-sovereign from the timer interrupts from the untrusted hypervisor with its synchronous tick delivery. This allows IncognitOS to reliably monitor the frequency of the hypervisor's possession of execution control (i.e., VMExits) and adjust the frequency of memory rerandomization performed by the paging subsystem, which transparently performs memory rerandomization through direct MMU access. The resulting IncognitOS design makes a case for self-obfuscating unikernel as a secure CVM deployment strategy while further advancing the obfuscation technique compared to previous works. Evaluation results demonstrate IncognitOS's resilience against CVM attacks and show that its adaptive obfuscation scheme enables practical performance for real-world programs.
-
- 작성일 2025-03-28
- 조회수 491
-
- [25.03.04] One paper accepted at EuroS&P 2025 from Professor Simon S Woo's (DASH Lab)
- The Data-driven AI & Security HCI Lab (DASH Lab, Advisor: Simon S. Woo) has had one System of Knowledge (SoK) paper accepted for publication at the 10th IEEE European Symposium on Security and Privacy (Euro S&P), a prestigious international conference covers Machine Learning Security, System & Network Security, Cryptographic Protocols, Data Privacy. The papers will be presented in July in Venice, Italy. SoK: Systematization and Benchmarking of Deepfake Detectors in a Unified Framework, EuroS&P 2025 Authors: Binh Le and Jiwon Kim (Ph.D. Student, Department of Software), Simon S. Woo (Professor, Department of Software, Sungkyunkwan University) This work is jointly performed with CSIRO Data61 as an international collaboration. Paper Link: https://arxiv.org/abs/2401.04364
-
- 작성일 2025-03-04
- 조회수 562
-
- [25.03.04] 2024 SKKU Fellowship 10 Professors Selected(Professor Hyungsik Kim of the College of Software Convergence)
- 2024 SKKU Fellowship 10 Professors Selected Our university has selected Professor Minwoo Kim of the College of Social Sciences, Professor Ahyoung Seo of the College of Business Administration, Professor Donghee Son of the College of Information and Communications, Professor Jaehyuk Choi of the College of Information and Communications, Professor Hyungsik Kim of the College of Software Convergence, Professor Jooyoung Shin of the College of Pharmacy, Professor Jaeyeol Cho of the College of Life Sciences, Professor Sehoon Lee of the College of Medicine, Professor Youngmin Kim of the Sungkyunkwan University Convergence Institute, and Professor Honghee Won of the Samsung Institute of Convergence Medicine and Science as the '2024 SKKU-Fellowship' professors. The SKKU-Fellowship system is the highest honor that our university has awarded since 2004, and is a system that selects the best professors whose research capabilities or industry-university cooperation achievements have reached world-class standards or are highly accessible, and grants them exceptional research support and honor. The 2024 SKKU-Fellowship is based on the university management policy of “Inspiring Future, Grand Challenge” for the 23rd and 24th academic years, and selected recipients in the fields of renowned international conferences, top journals and papers, and industry-academia cooperation ecosystems by expanding the excellence and scope of each professor. The awards ceremony was held at the general faculty meeting held on Wednesday, February 19th last year. This year, the 2024 SKKU Fellowship officially recommended candidates through the Fellowship Advisory Board, which advised on the selection of candidates, and Director of Industry-Academia Cooperation Gu Ja-chun, a member of the Fellowship Advisory Board, personally announced the 10 recipients. ▲ (From the top left) Professor Kim Min-woo, Professor Seo Ah-young, Professor Son Dong-hee, Professor Choi Jae-hyeok, Professor Kim Hyung-sik, Professor Shin Joo-young, Professor Cho Jae-yeol, Professor Kim Young-min Professor Kim Min-woo of the College of Social Sciences and Professor Kim Young-min of the Sungkyunkwan University of Convergence gave their acceptance speeches on behalf of the awardees. In the future, our university will continue to discover the various excellent achievements and values of the best professors and move forward to become a first-class university that contributes to human society.
-
- 작성일 2025-03-04
- 조회수 585
-
- [25.02.11] Professor Woo Hongwook’s Research Lab (CSI Lab), ICLR 2025 Paper Acceptance
- [Professor Woo Hongwook’s Research Lab (CSI Lab), ICLR 2025 Paper Acceptance] Two papers from CSI Lab (Supervised by Professor Woo Hongwook) have been accepted for presentation at ICLR 2025 (The 13th International Conference on Learning Representations), a prestigious conference in the field of Artificial Intelligence. The papers will be presented in April 2025 at the Singapore Expo in Singapore. 1. Paper “Model Risk-sensitive Offline Reinforcement Learning” The author of this paper is Kwangpyo Yoo, a Ph.D. candidate in the Department of Software. This study proposes a Model Risk-sensitive Reinforcement Learning (Model Risk-sensitive RL) framework for critical mission domains, such as robotics and finance, where decision-making is crucial. The paper particularly details a model risk-sensitive offline reinforcement learning technique (MR-IQN). MR-IQN aims to minimize the "model risk" loss in cases where the model's learned data differs from the real environment, leading to decreased accuracy. To achieve this, it calculates the model's confidence in each data point and evaluates the model risk per data point using a Critic-Ensemble Criterion. It also introduces a Fourier Feature Network that limits the gap between the actual policy's value function and the inferred policy’s value in an offline setting. MR-IQN outperformed other state-of-the-art risk-sensitive reinforcement learning techniques in experiments conducted in MT-Sim (financial trading environment) and AirSim (autonomous driving simulator), achieving lower risk and higher average performance. 2. Paper “NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks In Open Domains” This paper was co-authored by Wonje Choi (Ph.D. candidate, Department of Software), Jinwoo Park (Master’s student, Department of Artificial Intelligence), Sanghyun Ahn (Master’s student, Department of Software), and Daehui Lee (Integrated Master’s and Ph.D. candidate). The study proposes a Neuro-symbolic Continual Learner (NeSyC) framework that continuously generalizes knowledge (Actionable Knowledge) from embodied experiences to be applied to various tasks in open-domain physical environments. NeSyC mimics the human cognitive process of hypothesizing and deducing (hypothetico-deductive reasoning) to improve performance in open domains. This is achieved by: Using LLMs and symbolic tools to repeatedly generate and verify hypotheses from acquired experiences in a contrastive generality improvement approach. Utilizing memory-based monitoring to detect action errors of embodied agents in real-time and refine their knowledge, ultimately improving the agent's task performance and generalization across open-domain environments. NeSyC was evaluated across various benchmark environments, including ALFWorld, VirtualHome, Minecraft, RLBench, and real-world robotic tabletop scenarios. It demonstrated robust performance across dynamic open-domain environments and outperformed state-of-the-art methods, such as AutoGen, ReAct, and CLMASP, in task success rates. CSI Lab conducts research on network and cloud system optimization, autonomous driving of robots and drones, and other self-learning technologies by leveraging Embodied Agent, Reinforcement Learning, and Self-Learning. Contact Information:Professor Woo Hongwook | hwoo@skku.edu | CSI Lab | https://sites.google.com/view/csi-agent-group
-
- 작성일 2025-02-11
- 조회수 807
-
- [25.02.07] 2024 SW Talent Festival Grand Prize RISE Team (CSE '23): Interview with Team Leader Jung Gi-yong
- [25.02.07] 2024 SW Talent Festival Grand Prize RISE Team (Department of Computer Science and Engineering '23): Interview with Team Leader Jung Gi-yong From left: Jeong Hee-seong, Jung Gi-yong, Lee Gyu-min, Lee Sang-yeop, Lee Sang-jun On December 5th and 6th, 2024, the 2024 SW Talent Festival was held over two days. The event was hosted by the Ministry of Science and ICT and organized by the Information and Communication Planning and Evaluation Institute and the SW-Centric University Council. Under the theme “An AI World Connected by Software,” the festival showcased, exhibited, and awarded the major achievements and outstanding outcomes from 58 SW-centric universities. At this festival, the RISE team—composed of five students from the Department of Computer Science and Engineering at Sungkyunkwan University—won the Grand Prize (Minister of Science and ICT Award) by enhancing chart recognition performance through the construction of a Korean chart learning dataset. Let’s meet team leader Jung Gi-yong from the RISE team. Q: What motivated you to participate in the 2024 SW Talent Festival? A: In our Department of Computer Science and Engineering, there is a one-year industry–academia collaboration project. Out of roughly 20 teams that complete this program, professors select the most promising teams through meetings and offer them the opportunity to represent the university in competitions. My professor suggested that we participate, and that’s how we entered the contest under the name “RISE” (which carries the meanings “to awaken” and “to soar”). Q: The RISE team won the Grand Prize with “ChartBrain.AI.” Could you please explain what ChartBrain.AI is? A: ChartBrain.AI is a compact AI model that converts chart images into table text. When we started the project in April, we noticed that although the GPT-4o model understood general photographs and images quite well, it struggled with chart images—it lacked the ability to accurately extract numerical data. To address this shortcoming, we set out to develop a small AI module that converts chart images into table text in a format that can be easily understood by a Large Language Model (LLM). Because cloud-based GPT-4o models carry the risk of data leakage, companies typically do not use them for processing internal reports and chart images that require high security; instead, they deploy their own in-house LLMs. Our ChartBrain.AI is well suited for such applications. It is compact and, among domestic models, has achieved state-of-the-art (SOTA) accuracy at its current stage. Q: Could you describe the process of creating ChartBrain.AI? A: We began by taking an English model called Deplot—released by Google Research—and performed initial training to enable it to understand Korean. Then, we built a dataset of 1.12 million chart-to-table data pairs and further trained the Deplot model with this data to complete our system. Out of these, 320,000 pairs were synthetic chart images that our team created to supplement the limited diversity of existing open-source chart image datasets and to enable the model to understand even more complex charts. Note: LLM (Large Language Model): A language model built from neural networks with an enormous number of parameters. Deplot Official Code Link: https://github.com/google-research/google-research/tree/master/deplot Q: I heard that at your university’s booth you explained your award-winning work while wearing the Cheonggeumbok—the traditional attire once worn by Sungkyunkwan students. What prompted you to wear it, and what are your impressions? A: We didn’t prepare the Cheonggeumbok ourselves. The faculty member in charge of the Software Convergence College advised us to wear it during our presentation at the booth. Honestly, I felt a bit self-conscious walking around in it at first, but later, after seeing photos where it was immediately obvious that we were Sungkyunkwan University students, I grew to appreciate it. It seems our professors had remarkable foresight. Q: As the team leader, what was most important to you during the project? A: In an industry–academia collaboration project, the topic is provided by a company and the final deliverable must be submitted to them. This setup creates a greater sense of responsibility compared to typical school assignments—I constantly thought, “If we don’t deliver a proper result, it’ll be a big problem.” No matter how hard we worked, if the final program’s performance was poor, all our efforts would have been in vain. We had to keep pushing until we achieved the desired outcome. Although the process wasn’t easy, I believe that our relentless effort ultimately led us to produce excellent results. Q: What were some of the challenges you faced in submitting your program for the festival, and how did you overcome them? A: Everything was completely new to us. During the summer break, we participated in an “Industry–Academia Summer Intensive Work Program.” We rented a classroom and, much like interns, worked there from 9 a.m. to 6 p.m. every weekday. We dedicated ourselves to technical development and reading research papers. Through this process, we experienced for the first time the full cycle of reading papers, experimenting with prior research, setting a research direction based on experimental results, forming hypotheses, training models, reviewing results, and identifying shortcomings. As the team leader, I felt an even greater sense of responsibility. I believe our advisor’s active guidance and the enthusiastic participation of all team members were key to our success. Without our advisor, it would have been extremely difficult to succeed with this project. Q: Were there any particularly memorable moments during the project’s progress? A: We did not win a major award from the start. Before receiving the Grand Prize, we participated in two other competitions but were eliminated in the first round in both cases. When we entered the first competition, everyone worked incredibly hard—even staying up until 1 a.m. (the dormitory curfew) to continue development. For the second competition, we made further improvements over our previous version. By the time we entered this final competition, our morale was quite low compared to the first attempt; however, it’s very gratifying that our work eventually shone through. Q: You entered several competitions with the same project. Did the performance of the program improve significantly over time? A: Yes, there were significant improvements in performance. As mentioned earlier, our benchmark was the GPT-4 Omni model. While both our ChartBrain.AI and GPT-4 Omni were evolving simultaneously, during the summer break we were confident in stating that our model was superior. However, by December GPT-4 Omni had caught up with us for a period. In the end, our model advanced considerably and regained a comparative edge. This improvement was crucial in helping us win the Grand Prize. Q: Since the RISE team is composed entirely of CSE students, were there any classes or extracurricular activities that helped with this project? A: For me, participating in academic societies was extremely helpful. I joined an AI society called “TNT” at our university and took part in paper study sessions. As a sophomore, I was able to read many research papers, review them, and ask questions in TNT, which taught me how to discern a good paper from a less effective one. Personally, the paper review sessions in TNT were the most beneficial. Q: What do you find most attractive about studying software? A: Ultimately, software is about programming to create documents. In any company, the work I do might involve editing just a word or two in countless documents—and yet, those small changes can have a tremendous impact. Since text can be easily reproduced, even a small idea that improves one part can have infinite influence. I think that is the greatest appeal of software and what continuously fuels my competitive spirit. Q: What are your future career goals or aspirations? A: In the short term, I plan to write an undergraduate paper based on our award-winning project around May or June next year. When I work on development, I feel that roughly 50% of the help comes from GPT and about 30% from other online sources, which sometimes leaves me with the impression that I haven’t fully grasped everything. Therefore, my long-term goal is to pursue graduate studies so that I can deepen my knowledge in mathematics, English, and algorithms. Q: Do you have any advice for students preparing for software-related competitions? A: I believe that students preparing for competitions are already incredibly diligent, so here’s a tip rather than just advice. As you work on your project, you might find yourself deeply attached to your ideas. Explaining these ideas to others, especially to the judges, can be very challenging. This was the aspect that hindered me the most during competitions. I spent a lot of time thinking about how to effectively communicate my beloved idea, particularly to the judges. In my experience, effective persuasion can be achieved within just one or two PowerPoint slides. Focus on that key part, and I hope you achieve excellent results. Convince everyone possible, and I wish you great success.
-
- 작성일 2025-02-07
- 조회수 702
-
- [25.01.23] IIS Lab, Four Papers Accepted at NAACL 2025
- [25.01.23] IIS Lab, Four Papers Accepted at NAACL 2025 The Information and Intelligent Systems Lab (IIS Lab), led by Professor Ji-Hyung Lee, has had four papers accepted at NAACL 2025 (2025 Annual Conference of the North American Chapter of the Association for Computational Linguistics), one of the top-tier international conferences in natural language processing (NLP). The papers will be presented in April 2025 in New Mexico, USA. 1. DeCAP: Context-Aware Prompt Generation for Debiased Zero-shot Question Answering in Large Language Models (NAACL 2025) Authors: Sooyoung Bae (PhD Student, Department of Artificial Intelligence) Yoonseok Choi (Assistant Professor, Department of Computer Education, Sungkyunkwan University / PhD Graduate, Department of Software) Large language models (LLMs) perform well in zero-shot question answering (QA) tasks. However, existing methods suffer from performance gaps between ambiguous and clear questions and low debiasing performance due to strong dependence on provided instructions or internal knowledge. To address these issues, we propose DeCAP (Context-Aware Prompt Generation), which: Utilizes a Question Ambiguity Detector to reduce performance gaps caused by ambiguous question types. Employs a Neutral Next Sentence Generator to decrease dependency on internal biased knowledge by providing neutral contextual information. Experiments on BBQ and UNQOVER datasets across six LLMs show that DeCAP achieves state-of-the-art debiasing performance in QA tasks, significantly enhancing the fairness and accuracy of LLMs across diverse QA environments. 2. SALAD: Improving Robustness and Generalization through Contrastive Learning with Structure-Aware and LLM-Driven Augmented Data (NAACL 2025) Authors: Sooyoung Bae (PhD Student, Department of Artificial Intelligence) Hyojun Kim (SKT / MS Graduate, Department of Artificial Intelligence) Yoonseok Choi (Assistant Professor, Department of Computer Education, Sungkyunkwan University / PhD Graduate, Department of Software) This paper introduces SALAD (Structure-Aware and LLM-driven Augmented Data), a novel approach aimed at enhancing robustness and generalization in NLP models using contrastive learning. SALAD generates: Structure-aware positive samples using a tagging-based method. Counterfactual negative samples with diverse sentence patterns generated by LLMs. This allows the model to learn structural relationships between key sentence components while minimizing reliance on spurious correlations. We evaluate SALAD on three tasks: Sentiment Classification Sexism Detection Natural Language Inference (NLI) Results show that SALAD improves robustness and performance across different settings, including out-of-distribution datasets and cross-domain scenarios. 3. CoRAC: Integrating Selective API Document Retrieval with Question Semantic Intent for Code Question Answering (NAACL 2025) Authors: Yoonseok Choi (Assistant Professor, Department of Computer Education, Sungkyunkwan University / PhD Graduate, Department of Software) Cheolwon Na (Integrated MS/PhD Program, Department of Artificial Intelligence) Automated Code Question Answering (AQA) aims to generate precise answers for code-related queries by analyzing code snippets. However, in real-world settings, users often provide only partial code, making it difficult to derive correct answers. To address this challenge, we propose CoRAC, a knowledge-driven framework that improves AQA by: Selective API document retrieval Question semantic intent clustering We evaluate CoRAC on three real-world benchmark datasets, demonstrating its effectiveness. Results show that CoRAC generates high-quality answers outperforming LLM-based solutions like ChatGPT. 4. Q-FAKER: Query-free Hard Black-box Attack via Controlled Generation (NAACL Findings 2025) Authors: Cheolwon Na (Integrated MS/PhD Program, Department of Artificial Intelligence) Yoonseok Choi (Assistant Professor, Department of Computer Education, Sungkyunkwan University / PhD Graduate, Department of Software) Adversarial attack methods for testing language model vulnerabilities often require multiple queries and access to target model information. Even black-box attacks typically depend on target model output data, making them impractical in hard black-box settings where access is restricted. Existing hard black-box attack methods still demand high query counts and expensive adversarial generator training costs. To solve this, we introduce Q-FAKER (Query-free Hard Black-box Attacker), an efficient adversarial example generation method that: Uses a surrogate model to generate adversarial sentences without accessing the target model. Leverages controlled generation techniques for adversarial text generation. We evaluate Q-FAKER across eight datasets, demonstrating its high transferability and effectiveness in hard black-box attack scenarios. Contact Information Professor Ji-Hyung Lee | john@skku.edu IIS Lab | https://iislab.skku.edu/
-
- 작성일 2025-02-04
- 조회수 668