https://dspace.ewha.ac.kr/handle/2015.oak/267952 2026-04-05T00:31:13Z 2026-04-05T00:31:13Z 민동보 https://dspace.ewha.ac.kr/handle/2015.oak/274605 2026-03-23T16:30:03Z 2025-01-01T00:00:00Z

Title: UniTT-Stereo: Unified Training of Transformer for Enhanced Stereo Matching Ewha Authors: 민동보 Abstract: Unlike other vision tasks where Transformer-based approaches are becoming increasingly common, stereo depth estimation is still dominated by convolution-based models. This is mainly due to the limited availability of real-world ground truth for stereo matching, which hinders the performance improvement of transformer-based stereo approaches. In this paper, we propose UniTT-Stereo, a method to maximize the potential of Transformer-based stereo architectures by unifying self-supervised learning for pre-training with stereo matching framework based on supervised learning. Specifically, we design a dual-task learning scheme that reconstructs masked regions of an input image while simultaneously predicting corresponding points in the paired image. We demonstrate that this approach encourages the model to learn locality-aware representations, which are critical to overcoming the data inefficiency of Transformers. Moreover, to address these challenging tasks of reconstruction-and-prediction, we propose a variable masking ratio strategy that promotes robustness to varying levels of visual information. Additionally, we introduce losses that exploit stereo geometry and correspondence at the appearance, feature, and disparity levels. To further validate the effectiveness of our design, we conduct frequency decomposition and attention map visualization, which reveal how the model effectively captures fine-grained structures and cross-view correspondences. State-of-the-art performance of UniTT-Stereo is validated on various benchmarks such as the ETH3D, KITTI 2012, and KITTI 2015 datasets. Code is available at: https://github.com/00kim/UniTT-Stereo

2025-01-01T00:00:00Z 반효경 https://dspace.ewha.ac.kr/handle/2015.oak/273682 2026-01-14T16:31:04Z 2025-01-01T00:00:00Z

Title: Rethinking I/O Caching for Large Language Model Inference on Resource-Constrained Mobile Platforms Ewha Authors: 반효경 Abstract: Large language models (LLMs) have traditionally relegated inference to remote servers, leaving mobile devices as thin clients. Recently, advances in mobile GPUs and NPUs have made on-device inference increasingly feasible, particularly for privacy-sensitive and personalized applications. However, executing LLMs directly on resource-constrained devices exposes severe I/O bottlenecks, as repeated accesses to large weight files can overwhelm limited memory and storage bandwidth. Prior studies have focused on internal mechanisms such as KV caching, while the role of the host OS buffer cache remains underexplored. This paper closes that gap with file-level trace analysis of real-world mobile LLM applications, and identifies three characteristic access patterns: (1) one-time sequential scans during initialization, (2) persistent hot sets (e.g., tokenizers, metadata, indices), and (3) recurring loop accesses to model weight files. Guided by these observations, we propose LLM-aware buffer cache strategies and derive cache-sizing guidelines that relate loop size, host-set coverage, and storage bandwidth. We further compare smartwatch-class and smartphone-class platforms to clarify feasible model sizes and practical hardware prerequisites for local inference. Our results provide system-level guidance for I/O subsystem design that enables practical on-device LLM inference in future mobile and IoT devices.

2025-01-01T00:00:00Z 김영준 https://dspace.ewha.ac.kr/handle/2015.oak/273535 2026-01-14T05:16:18Z 2025-01-01T00:00:00Z

Title: Virtual Reenactment of the Itaewon Crowd Crush Using Kinodynamic Simulation Ewha Authors: 김영준 Abstract: We propose new crowd simulation methods to virtually reenact the Itaewon disaster that occurred in 2022 due to the extreme crowd density. Conventional techniques make it challenging to simulate diverse, extremely dense crowd behaviors such as crowd surge, fluidization, and falls observed at the Itaewon disaster. This paper proposes a kinodynamic agent simulation combining kinematic agents for low-density crowds, hydrodynamic and hydrostatic agents for high-density crowds, and articulated passive agents for high-density crowds in dense contact. In order to perform co-simulation among heterogeneous agent types, we use a message-passing mechanism to share relevant kinematic and dynamic information among agents and make agent-type transitions based on crowd density and contact forces. Experiments show that the proposed hybrid simulation approach can accurately reenact crowd phenomena observed at the Itaewon compared to the CCTV footage. Moreover, our ablation study supports the use of kinodynamic agents to faithfully reconstruct the Itaewon crowd behavior. Furthermore, we run three what-if scenarios to explore the possibilities of using our techniques to help prevent incidents in the future. Finally, to demonstrate the applicability of our proposed methods to other types of extreme crowd behaviors besides the Itaewon disaster, we simulate two other real-world crowd incidents using our techniques.

2025-01-01T00:00:00Z 도인실 https://dspace.ewha.ac.kr/handle/2015.oak/273411 2026-03-19T16:31:25Z 2026-01-01T00:00:00Z

Title: GIIDS-AR: End-to-end generalized intelligent intrusion detection system with adversarial robustness for heterogeneous UAVs in UAM Ewha Authors: 도인실 Abstract: This study presents GIIDS-AR, an enhanced version of the Generalized Intelligent Intrusion Detection System (GIIDS), developed to enhance robustness and secure diverse Unmanned Aerial Vehicles (UAVs) in Urban Air Mobility (UAM) while preserving generalization. As UAVs grow vital in logistics, emergency response, and disaster relief, their reliance on wireless communication increases exposure to cyber threats. GIIDS leverages machine learning for cross-platform detection but remains vulnerable to adversarial machine learning (AML) attacks. To assess this, GIIDS was tested under black-box, white-box, and transfer attacks. Accuracy dropped to 72 % under black-box and recall to 49.9 % under white-box settings. Adversarial training restored original performance improving accuracy to 99.0 % and F1 to 99.8 %, with AUC reaching 1.00. These evaluations were conducted using cross-dataset splits of live and simulated UAV telemetry, ensuring resilience on previously unseen data. GIIDS-AR retains layered modeling, time-aware feature encoding, and ensemble learning, while incorporating adversarial examples to improve resilience. It demonstrates strong detection performance under diverse attacks and generalizes effectively across heterogeneous UAV platforms. Our findings reveal that generalization techniques inherently contribute to adversarial robustness, positioning GIIDS-AR as one of the first unified UAV IDS frameworks capable of securing UAV networks against evolving cyber threats. © 2025 The Authors

2026-01-01T00:00:00Z