Hansung University

The widespread use of mobile devices has led to the continuous collection of vast amounts of user-generated data, supporting data-driven decisions across a variety of fields. However, the growing volume of these data raises significant privacy concerns, especially when they include personal information vulnerable to misuse. Differential privacy (DP) has emerged as a prominent solution to these concerns, enabling the collection of user-generated data for data-driven decision-making while protecting user privacy. Despite their strengths, existing DP-based data collection frameworks are often faced with a trade-off between the utility of the data and the computational overhead. To address these challenges, we propose the differentially private fractional coverage model (DPFCM), a DP-based framework that adaptively balances data utility and computational overhead according to the requirements of data-driven decisions. DPFCM introduces two parameters, α and β, which control the fractions of collected data elements and user data, respectively, to ensure both data diversity and representative user coverage. In addition, we propose two probability-based methods for effectively determining the minimum data each user should provide to satisfy the DPFCM requirements. Experimental results on real-world datasets validate the effectiveness of DPFCM, demonstrating its high data utility and computational efficiency, especially for applications requiring real-time decision-making.

As quantum computing progresses, the assessment of cryptographic algorithm resilience against quantum attack gains significance interests in the field of cryptanalysis. Consequently, this paper proposes the depth-optimized quantum circuit of Korean hash function (i.e., LSH) and estimates its quantum attack cost in quantum circuits. By utilizing an optimized quantum adder and employing parallelization techniques, the proposed quantum circuit achieves a 78.8% improvement in full depth and a 79.1% improvement in Toffoli depth compared to previous the-state-of art works. In conclusion, based on the proposed quantum circuit, we estimate the resources required for a Grover collision attack and evaluate the post-quantum security of LSH algorithms.

Background: Accurate determination of the natural head position (NHP) is essential in orthognathic surgery for optimal surgical planning and improved patient outcomes. However, traditional methods encounter reproducibility issues and rely on external devices or patient cooperation, potentially leading to inaccuracies in the surgical plan. Methods: To address these limitations, we developed a geometric deep learning network (NHP-Net) to automatically reproduce NHP from CT scans. A dataset of 150 orthognathic surgery patients was utilized. Three-dimensional skull meshes were converted into point clouds and normalized to fit within a unit sphere. NHP-Net was trained to predict a 3 × 3 rotation matrix to align the CT-acquired posture with the NHP. Experiments were conducted to determine optimal point cloud sizes and loss functions. Performance was evaluated using mean absolute error (MAE) for roll, pitch, and yaw angles, as well as a rotation error (RE) metric. Results: NHP-Net achieved the lowest RE of 1.918° ± 1.099° and demonstrated significantly lower MAEs in roll and pitch angles compared to other deep learning models (p < 0.05). These findings indicate that NHP-Net can accurately align CT-acquired postures to the NHP, enhancing the precision of surgical planning. Conclusions: By effectively improving the accuracy and efficiency of NHP reproduction, NHP-Net reduces the workload of surgeons, supports more precise orthognathic surgical interventions, and ultimately contributes to better patient outcomes.

In this paper, we propose a hybrid genetic algorithm (HGA) that embeds the tabu search mechanism into the genetic algorithm (GA) for multiple-input multiple-output (MIMO) detection. We modified the selection and crossover operation to maintain the diverse and wide exploration areas, which is an advantage of the GA, and the mutation operation to perform a local search for a specific region. In the mutation process, the ’tabu’ concept is also employed to prevent the repeated search of the same area. In addition, a layered detection process is applied simultaneously with the proposed algorithm, which not only improves the bit error rate performance but also reduces the computational complexity. We apply the layered HGA (LHGA) to the MIMO system with very high modulation order such as 64-quadrature amplitude modulation (QAM), 256-QAM, and 1024-QAM. Simulation results show that the LHGA outperforms conventional detection approaches. Especially, in the 1024-QAM MIMO system, the LHGA has less than 10% of computational complexity but a 6 dB signal-to-noise ratio (SNR) gain compared to the conventional GA-based MIMO detection scheme.

Background: The large language model (LLM) has the potential to be applied to clinical practice. However, there has been scarce study on this in the field of gastroenterology. Aim: This study explores the potential clinical utility of two LLMs in the field of gastroenterology: a customized GPT model and a conventional GPT-4o, an advanced LLM capable of retrieval-augmented generation (RAG). Method: We established a customized GPT with the BM25 algorithm using Open AI’s GPT-4o model, which allows it to produce responses in the context of specific documents including textbooks of internal medicine (in English) and gastroenterology (in Korean). Also, we prepared a conventional ChatGPT 4o (accessed on 16 October 2024) access. The benchmark (written in Korean) consisted of 15 clinical questions developed by four clinical experts, representing typical questions for medical students. The two LLMs, a gastroenterology fellow, and an expert gastroenterologist were tested to assess their performance. Results: While the customized LLM correctly answered 8 out of 15 questions, the fellow answered 10 correctly. When the standardized Korean medical terms were replaced with English terminology, the LLM’s performance improved, answering two additional knowledge-based questions correctly, matching the fellow’s score. However, judgment-based questions remained a challenge for the model. Even with the implementation of ‘Chain of Thought’ prompt engineering, the customized GPT did not achieve improved reasoning. Conventional GPT-4o achieved the highest score among the AI models (14/15). Although both models performed slightly below the expert gastroenterologist’s level (15/15), they show promising potential for clinical applications (scores comparable with or higher than that of the gastroenterology fellow). Conclusions: LLMs could be utilized to assist with specialized tasks such as patient counseling. However, RAG capabilities by enabling real-time retrieval of external data not included in the training dataset, appear essential for managing complex, specialized content, and clinician oversight will remain crucial to ensure safe and effective use in clinical practice.

This study conducted multi-sensor image classification by utilizing Google Earth Engine (GEE) and applying satellite imagery from Korean Multi-purpose Satellite 3 (KOMPSAT-3), KOMPSAT-5 SAR, Compact Advanced Satellite 500-1 (CAS500-1), Sentinel-1, and Sentinel-2 within GEE. KOMPSAT-3/5 and CAS500-1 images are not provided by GEE. The land-use and land-cover (LULC) classification was performed using the random forest (RF) algorithm provided by GEE. The study experimented with 10 cases of various combinations of input data, integrating Sentinel-1/-2 imagery and high-resolution imagery from external sources not provided by GEE and those normalized difference vegetation index (NDVI) data. The study area is Boryeong city, located on the west coast of Korea. The classified objects were set to six categories, reflecting the region’s characteristics. The accuracy of the classification results was evaluated using overall accuracy (OA), the kappa coefficient, and the F1 score of the classified objects. The experimental results show a continued improvement in accuracy as the number of applied satellite images increased. The classification result using CAS500-1, Sentinel-1/-2, KOMPSAT-3/5, NDVI from CAS500-1, and NDVI from KOMPSAT-3 achieved the highest accuracy. This study confirmed that the use of multi-sensor data could improve classification accuracy, and the high-resolution characteristics of images from external sources are expected to enable more detailed analysis within GEE.

Domain adaptation aims to effectively learn from two domains with different distributions, solving labeling problems; however, traditional methods assume that the source and target data are in-distribution data that share the same labels. In practice, Out-Of-Distribution (OOD) data which do not share labels with the existing data may also be collected during the target data collection process. These OOD data introduce noise and confusion, leading to decreased performance during adaptation. To address this issue, we propose RObust Domain Adaptation from Out-Of-Distribution data (RODA-OOD), a novel method based on data-centric AI principles that focuses on improving data quality rather than refining model architecture. RODA-OOD utilizes the characteristics of deep learning models that prioritize learning in-distribution data, which are easier to train on compared to OOD data. By dynamically adjusting the threshold for OOD detection, the proposed method effectively filters out OOD data, allowing the model to focus on relevant target data. RODA-OOD was compared with competitor and original domain adaptation algorithms based on target data accuracy. The results show that RODA-OOD demonstrates the most robust performance against OOD data, achieving a 21.3% increase in accuracy compared to existing domain adaptation methods. Thus, RODA-OOD can provide a solution to the OOD issue in unsupervised domain adaptation.

This paper addresses the rendezvous problem in asymmetric cognitive radio networks (CRNs) by proposing a novel reinforcement learning (RL)-based channel-hopping algorithm. Traditional methods like the jump-stay (JS) algorithm, while effective, often struggle with high time-to-rendezvous (TTR) in asymmetric scenarios where secondary users (SUs) have varying channel availability. Our proposed RL-based algorithm leverages the actor-critic policy gradient method to learn optimal channel selection strategies by dynamically adapting to the environment and minimizing TTR. Extensive simulations demonstrate that the RL-based algorithm significantly reduces the expected TTR (ETTR) compared to the JS algorithm, particularly in asymmetric scenarios where M-sequence-based approaches are less effective. This suggests that RL-based approaches not only offer robustness in asymmetric environments but also provide a promising alternative in more predictable settings.