Driven by the challenges associated with training large models in heterogeneous networks, where a one-size-fits-all training strategy can result in inefficiencies, we propose a versatile Federated Split Learning (FSL) algorithm designed for dynamic resource allocation. Conventional methods often encounter resource limitations, particularly in settings with varying computational abilities. Our algorithm overcomes this issue by intelligently distributing model components among nodes according to their specific computational capabilities, thereby enhancing resource utilization.A key feature of our approach is the inclusion of an auxiliary classifier layer that enables each client to calculate losses and locally update parameters, while only sending activation values to the central server. This mechanism not only reduces communication overhead but also speeds up the overall training process. Additionally, we implement an earlyexit inference strategy that adapts based on task complexity and available resources, further enhancing inference efficiency.
This work addresses a route planning problem constrained by a bus road network that includes the schedules of all buses. Given a query with a starting bus stop and a set of Points of Interest (POIs) to visit, our goal is to find an optimal route on the bus network that allows the user to visit all specified POIs from the starting stop with minimal travel time, which includes both bus travel time and waiting time at bus stops. Although this problem resembles a variant of the Traveling Salesman Problem, it cannot be effectively solved using existing solutions due to the complex nature of bus networks, particularly the constantly changing bus travel times and user waiting times. In this paper, we first propose a modified graph structure to represent the bus network, accommodating the varying bus travel times and their arrival schedules at each stop. Initially, we suggest a brute-force exploration algorithm based on the Dijkstra principle to evaluate all potential routes and determine the best one; however, this approach is too costly for large bus networks. To address this, we introduce the EA-Star algorithm, which focuses on computing the shortest route for promising POI visit sequences. The algorithm includes a terminal condition that halts evaluation once the optimal route is identified, avoiding the need to evaluate all possible POI sequences. During the computation of the shortest route for each POI visiting sequence, it employs the $A^{*}$ algorithm on the modified graph structure, narrowing the search space toward the destination and improving search efficiency. Experiments using New York bus network datasets demonstrate the effectiveness of our approach.
Compute Express Link (CXL), a cache-coherent interconnect protocol for CPUs and PCIe devices, has emerged as a promising solution for memory expansion. In this work, we develop a CXL simulator built upon GEM5 and QEMU to assess the system-wide performance impacts based on varying attributes of CXL devices, such as memory capacity and latency. The simulator directs memory traffic to either system memory or the CXL memory expander, tracks cache hit/miss ratios, and projects application performance. Our simulation study demonstrates that the simulator accurately predicts application performance.
Multiple Object Tracking (MOT) has applications across many fields, yet occlusion-related issues continue to impact its effectiveness. These occlusions lead to information loss, reducing model performance and causing re-identification errors. In response, this study introduces a Dual View Sharing Multi Object Tracking (DVSMOT) method. Results show that this approach minimizes information loss due to occlusion, improves re-identification accuracy, and enhances tracking precision in experimental trials compared to traditional methods.
The rapid growth of the internet and digital technologies has expanded access to digital content while increasing copyright infringement through illegal content-sharing websites. These platforms cause significant financial losses to copyright holders and persist despite government monitoring efforts. This study proposes an automated method to analyze piracy site characteristics and classify sites involved in copyright infringement. By addressing the evolving nature of such platforms, the approach enhances detection and contributes to combating online piracy effectively.
To build a model that predicts the gender of named entities in text, a high-quality labeled dataset is required, which requires considerable manual effort and time. This paper proposes two major contributions to address this issue. First, we develop a mechanism to automatically extract individual entities from sentences using Named Entity Recognition (NER) and Part-of-Speech (PoS) tagging techniques. This approach automates the data generation process and reduces costs. Second, we ensure class balance in the dataset to optimize model performance. Experimental results demonstrate that the automated data generation method and balanced dataset significantly enhance the performance of the gender prediction model. This work makes a substantial contribution to data generation and the improvement of model performance in Natural Language Processing (NLP) tasks.
Retrieval-Augmented Generation (RAG) is aimed at improving the functionality of large language model (LLM) applications by incorporating specific data. This may include searching for relevant materials or files concerning a particular issue or search provided as background information to the LLM. Some prominent architectures like the Vanilla RAG architecture focuses primarily on retrieval of textual data, primarily utilizing vector databases, thus neglecting the structural intricacies of textual data, resulting in a critical gap in the generation process. To address this gap, we have introduced SAGE-QA (Scalable Advanced Graph RAG for Enterprise Question Answering), which significantly enhances both the retrieval and generation processes by emphasizing the importance of topological information for reasoning tasks on a variety of data sources. For our use case, SAGE-QA significantly outperforms current state-of-the-art RAG methods with its novel retrieval mechanism, which effectively mitigates hallucinations, and acts as a sustainable contextually relevant application.
Table question answering datasets have been released in various table sources and diverse tasks. However, most of these table question answering datasets are constructed in resource-rich languages like English. For low-resource languages, it is necessary to create new datasets or translate existing ones to train models on table question answering tasks. This process requires substantial costs and manpower. Table data, unlike natural text, is structured in a way that is less affected by word order and language-specific characteristics. In this paper, we propose a method to minimize the costs associated with translating and applying new table question answering datasets. By leveraging the characteristics of table data, we suggest translating only the question data of the question answering table datasets into multiple languages. To minimize the performance degradation caused by using the native table data of the original dataset and not directly utilizing the table data in the target language, we applied adaptation methods using contrastive learning and adversarial training. To enhance the reasoning capabilities required for table question answering without diminishing performance in languages other than the one used for pre-training, we further pre-trained a RoBERTa-based multilingual encoder on table data using SQL data. The proposed cross-lingual adaptation demonstrated performance improvements in most of the languages evaluated. Additionally, the pre-training using SQL data significantly improved performance across all languages, even though the pre-training table data consisted of English data.
As the field of robotics advances, Embodied Instruction Following (EIF) has emerged as a key challenge in artificial intelligence. EIF tasks require agents to interpret and execute natural language instructions by predicting and completing sequences of subgoals within a physical environment. Traditional evaluation metrics—such as Success Rate (SR), Goal Condition (GC), Path Length Weighted SR (PLWSR), and Path Length Weighted GC (PLWGC)—primarily focus on task success following low-level control actions. However, these metrics inadequately assess the accuracy of high-level planning, which is critical for overall task performance. Existing methods for evaluating high-level planning often rely on comparing predicted plans to a single human-annotated ground truth trajectory, implicitly assuming the existence of only one correct solution. In practice, many instructions allow for multiple valid trajectories that can achieve the same goal. To address these limitations, we propose Relaxed HLP, a novel metric designed to evaluate high-level planning more flexibly by accounting for alternative valid plans. Relaxed HLP introduces three key considerations: temporal agnosticism, spatial agnosticism, and interchangeable actions, thus enabling a more comprehensive assessment of high-level plan accuracy. We validate the effectiveness of Relaxed HLP through human evaluations, demonstrating that it aligns more closely with human judgment compared to traditional ground truth-based metrics. Our results underscore the robustness of Relaxed HLP in capturing diverse, semantically equivalent plans, offering a more accurate assessment of high-level planning in EIF tasks.
With the increasing use of smartphones for accessing e-commerce (EC) websites, the need for supplementary post-login continuous authentication has grown, which results in the need for touch stroke datasets for research. Existing datasets for touch-based authentication, primarily focused on scenarios like reading text or social media, exhibit a directional bias in strokes, limiting their applicability in EC environments where touch interactions involve a wider variety of stroke directions. To address this, we introduce eCommTouch, a new public dataset that simulates touch interactions on EC websites. The eCommTouch dataset includes a diverse set of strokes in leftward, rightward, upward, and downward directions, and provides a larger number of strokes per session than existing datasets. Our preliminary evaluation using equal error rate (EER) on the eCommTouch dataset revealed that stroke direction-specific classifiers enhance authentication performance, i.e., lowering EER, when a user has nearly balanced strokes in each direction, such as leftward and rightward, which has not been confirmed with existing datasets. We anticipate that the eCommTouch dataset will contribute to advancing touch-based authentication, particularly in enhancing security for EC platforms by leveraging diverse stroke patterns. The eCommTouch dataset is available from https://github.com/yamanalab/eCommTouch.
Rising sea levels due to climate change increasingly threaten medical infrastructure through flooding. This study develops machine learning models to predict flood exposure for 11,508 medical facilities in the southeastern coastal regions of the United States by integrating datasets including meteorological, hydrological, topographic, and geological data, the Natural Risk Index, and historical flood records from NASA, HIFLD, and FEMA. Six regression models, namely Linear Regression, Support Vector Regression, Random Forest, k-Nearest Neighbors, XGBoost, and Artificial Neural Networks, are trained using 16 explanatory variables identified through literature review and correlation analysis. Data preprocessing employs the SMOGN for class imbalance and Winsorization for outliers. Model performance is evaluated using MAE, MSE, and RMSE, with Random Forest and XGBoost models achieving the highest performance (MSE of 2.58e-5 and 3.69e-5, respectively). This multifactorial approach allows the models to capture complex flood-influencing relationships, enhancing adaptability and performance across geographic regions. Future work focuses on expanding across the U.S. and developing a near real-time flood monitoring system.
Lie detection has been a critical area of research with applications in various fields such as forensics, criminal investigation, and psychology. Traditional methods like the polygraph have limitations, such as low accuracy, subjective interpretation, and invasive procedures, which hinder their practical use. Consequently, several studies have explored micro-expression analysis and voice signal-based detection, which show significant potential for detecting the authenticity of a conversation. However, these approaches face challenges in achieving high accuracy and generalizability. To address these issues, this paper proposes a multi-modal model-based lie detection system that integrates audio, text, and video features extracted from a single video. The system utilizes advanced algorithms for audio feature extraction and visual feature analysis. Additionally, a Vision-Language Model for multimodal data integration is employed to detect the authenticity of a conversation. This approach enables the robust detection of subtle signal changes in voice and visual expressions, providing a non-invasive and efficient solution. The proposed system is validated by demonstrating its effectiveness and potential through an actual dataset.
With the advent of smart meters in recent years, privacy has become one of the biggest concerns in smart grids. While privacy is achieved by designing smart grid services while maintaining the anonymous participation of consumers, several technical issues must be solved by electricity companies, such as confirmation of payments or assurance of consumption data integrity and quality. In this paper, we propose a new system model for data aggregation in smart grid environments that balances the needs of consumers and electricity companies. To decentralize and achieve transparent data processing, we apply a permissioned blockchain network and smart contracts. To support consumer privacy, we propose a privacy-preserving smart contract that runs on a trusted execution environment. Payments are processed and confirmed at the end of a billing period by the smart contract to enforce data consistency and integrity. We implement our smart contract design for Hyperledger Fabric, a permissioned blockchain network, and analyze its privacy, data, and blockchain benefits.
Software Defect Prediction (SDP) is crucial for ensuring the quality of software systems. While traditional and emerging transformer-based models are well researched in SDP, recent advancement of state space model - Mamba, has gained popularity in various domains. This research explores the potential of state space model in the SDP domain for efficiently extracting effective representations from data. Inspired by Mamba Tab's lightweight, scalable, and generalizable nature, we experimented to evaluate its performance in the context of SDP. Our experiment involved several datasets and compared Mamba Tab with traditional machine learning, state-of-the-art deep learning, and transformer - based models. The experimental results demonstrate that Mamba Tab outperforms other baseline models across most key metrics and time complexity analysis, further confirming its efficiency. Cohen's d effect size analysis strengthens this advantage, showing large and medium effect sizes for Mamba Tab on these metrics. These findings highlight Mamba Tab's effectiveness, efficiency and generalizability, in the context of SDP.
The computation and storage capabilities of a cloud-server can bring the benefits of high scalability and availability to an image retrieval application; however, their main limitation comes from the associated privacy concerns during data transmission, storage and computation. To ensure privacy, this study implements a content-based image retrieval application for finding different types of brain tumors in the encrypted domain. In this framework, we propose a perceptual encryption technique to protect images in a way that their features necessary for high-dimensional representation can still be extracted from the cipher-images. Experimental results show that compared to conventional secure techniques our proposed system reduces difference in non-secure and secure retrieval performance up to 3%.
Although text-to-image models excellently create realistic images from text, they struggle with long-form text due to token limits in the pretrained text encoder. In this paper, we propose the Padding Is Enough(PIE) text encoder that is trained to represent long-form text contexts with a single embedding. The embedding representing long-form text utilizes a knowledge distillation technique, where the outputs from the PIE text encoder and the CLIP text encoder are input into a diffusion model to align their outputs. Specifically, we dont train the diffusion model, but only the text encoder, thereby preserving the extensive pretrained knowledge. Furthermore, the PIE text encoder can be used to extend text prompt in task-specific large pretrained diffusion models. It helps enhance the expressiveness of large pretrained models while reducing costs. We demonstrate that our model achieves high availability as a text extension without damage to the extensive knowledge of the pretrained diffusion models.
Software Defect Prediction (SDP) ensures software quality by identifying defects early in development. However, a major challenge is class imbalance, where defect data is vastly outnumbered by non-defect data, reducing prediction accuracy. To address this, we propose DefectGRANDE, a hybrid model leveraging SMOTE for data balancing, RandomForest for feature importance, and ensemble learning for stability. DefectGRANDE outperforms existing methods in metrics like Positive Detection (PD), Balance, and AUC. Through this, defects can be predicted more effectively in the early stages of development, enhancing software quality and reliability while reducing development costs and time.
With the rapid expansion of data markets, accurately assessing the value of data has become critical to fostering trust and fairness in data transactions. However, the lack of standardized methods for evaluating the worth of personal information, particularly personal image data, poses a significant challenge. This study examines how different levels of de-identification impact the perceived value of image data. Using the Structural Similarity Index Measure (SSIM), we assess changes in image quality, and through regression analysis, we observe a marked decrease in data value as de-identification levels increase. Our findings emphasize the importance of developing compensation models that align privacy preservation with data valuation, contributing to the creation of a transparent and secure data trading environment.
Visual analytics are widely used across various domains, including business intelligence, healthcare, and e-commerce, to provide data-driven insights that enhance decision-making processes. These systems analyze large datasets to generate visualizations that help users identify trends, patterns, and actionable insights. These insights rely heavily on accurate and complete data. However, effectively addressing the challenges posed by missing data is crucial for ensuring the accuracy and reliability of data-driven insights and visualizations. Incomplete datasets often lead to reduced accuracy in visualizations, prompting users to adopt costly imputation strategies to clean the data. This paper presents an innovative approach, ‘CoDAQ’-Congressional-Based Data Imputation for Aggregate Queries that explores the use of a stratified-based selection technique for imputation to handle missing values within a specified imputation budget effectively. CoDAQ strategically selects cells for imputation by analyzing missing data patterns across different groups, ensuring a structured and representative approach. This technique offers enhanced accuracy and consistency over baseline methods. Our findings illustrate how the proposed technique enhances the utilization of available data, improves the accuracy of visualizations, and reduces bias in analysis, ultimately increasing the performance of visual analytics.
Improving the loss function can significantly enhance learning outcomes by more effectively addressing the differences between the model's predicted values and the actual values. Categorical Cross Entropy (CCE) calculates the loss using the negative logarithm of probability predictions, yielding smaller loss values for predictions closer to the correct answer and larger loss values for incorrect predictions. In deep learning, the model uses the gradient of the loss function to determine how much to update its weights. Once the model's predictions are sufficiently close to the correct answer, further weight updates are unnecessary. However, CCE has a limitation: its gradient is always less than -1, causing continuous weight updates. This results in overtraining for high-confidence predictions and undertraining for low-confidence predictions. When training on the Caltech-UCSD Birds 200 dataset, we observed that the original CCE loss function led to predicted values clustering near 1, with few small prediction values appearing as outliers. To address this, we propose a new loss function, Smooth CCE, which adjusts the gradient to approach zero as the predicted values approach 1. This prevents unnecessary updates and leads to a more balanced distribution of predictions, increasing the proportion of lower predicted values. Using Smooth CCE with lambda = 0.7 and lambda = 0.5, we observed performance improvements of 0.6%p on Efficient-NetV2B0, 2.7%p on MobileNetV3Small, and 2.0%p on InceptionV3 compared to the original CCE.
