Kangnam University

[This corrects the article DOI: 10.3389/fpsyg.2026.1731891.].

Containers have become a critical component of cloud-native technologies, enabling organizations to run scalable and isolated workloads. However, in recent years there has been an increase in the sophistication of attacks targeting the cloud-native environment. It is crucial to implement security controls that protect containers at all stages of their lifecycle to mitigate risks such as privilege escalation by malicious applications. Numerous studies have aimed to develop security profiles for container applications that limit their privileges and protect the host system from compromise. Despite these efforts, the shared underlying kernel may still permit several successful attacks. In this study, we strive to develop a concise system call whitelist to address the problem of excessive privileges while ensuring the operational availability of applications. To achieve this, we propose to enhance static analysis with dynamic analysis to gather comprehensive information about the containerized application during two distinct execution phases: the initialization and the serving phases. Using this information, we determine the essential system calls for the application's operation and prevent all unwarranted system calls. We then perform crash analysis on the container under test to identify and incorporate any missing system calls. Through numerous experiments with popular server applications, we confirm that our approach is effective in discovering the necessary system calls for the operation of containerized applications. The system call whitelists produced by this method are more concise than Docker's default seccomp profile, consequently reducing the attack surface for a wide variety of applications and host systems significantly.

In low-light environments, the visual information in images becomes limited, which often leads to a significant degradation in image classification performance. This is particularly problematic for transformer-based models such as the Vision Transformer, which relies on global attention learned from image patches and is therefore highly sensitive to input quality. In this study, we propose a preprocessing-based approach to improve the classification performance of Vision Transformer under low-light conditions. Specifically, we apply histogram equalization to input images in order to restore contrast and enhance visual features that are often lost in low-light scenes. This enhancement enables Vision Transformer to learn richer and more meaningful visual representations, resulting in improved classification accuracy. Experiments were conducted using the ExDark dataset, which contains images captured under real-world low-light conditions, and the effectiveness of the proposed preprocessing technique was quantitatively evaluated. The results demonstrate that even without modifying the model architecture, histogram-based preprocessing significantly improves the classification performance of Vision Transformer. This study highlights the importance of preprocessing in low-light image classification and presents an efficient strategy for enhancing the practical utility of transformer-based vision models.

In recent deep learning-based image classification studies, Vision Transformer has shown high performance and has the potential to replace CNN. In this study, we propose a new structure that integrates frequency domain information based on Fourier Transform into the ViT structure. The proposed model effectively combines spatial-frequency information of images while maintaining the global representation learning ability of the self-attention mechanism, thereby enabling more sophisticated visual feature representation. The training and validation loss curves showed that the proposed model showed an overall stable learning process, had little overfitting, and converged to the lowest validation loss. In the class-by-class accuracy evaluation, the highest accuracy was recorded in the Automobile, Deer, Horse, and Ship classes, and in particular, the Ship class showed a significant performance improvement compared to other models. This suggests that frequency-based information is effective for recognizing objects with repetitive structures or distinct outlines. On the other hand, in the Dog and Cat classes with complex backgrounds or ambiguous shapes, it showed a slight difference, but overall, it maintained a balanced performance. the proposed model showed superior performance in learning stability and classification accuracy compared to the existing ViT, and showed a particularly strong effect in tasks that require simultaneous consideration of object shape information and periodic patterns. Proposed frequency-based Attention mechanism suggests the possibility of expanding the Transformer series vision model, and is expected to be effectively utilized in various visual recognition problems such as medical imaging, satellite imaging, autonomous driving, and multimodal learning. In future studies, it can be expanded into a model applicable to real-time environments through fusion with various frequency transformation techniques and application of lightweight structures.

The VR-based circular tracking movement evaluation system (CES) was developed to quantitatively assess visuomotor control. The virtual stick, a component of the CES, provides visual cues in the virtual environment and haptic feedback when holding the controller. This study examined the effects of stick presence and presentation order on control accuracy for distance, angle, and angular velocity. Twenty-seven participants (12 females; mean age 23.3 ± 2.3 years) performed tasks in the frontal plane followed by the sagittal plane. In each plane, the stick was visible for the first 1-3 revolutions and invisible for the subsequent 4-6 revolutions in the invisible condition, with the reverse order in the visible condition. In the invisible condition, control accuracy with the stick was 1.10 times higher for distance only in the sagittal plane, and 1.13 and 1.09 times higher for angle and angular velocity in the frontal plane, and 1.11 and 1.08 times higher in the sagittal plane. No significant differences were observed in the visible condition. The improved control accuracy when the stick was visible is likely due to enhanced precision in constructing the reference frame, internal models, body coordinates, and effective multisensory integration of visual and haptic information. Such visual information may enable fine control in virtual environment-based applications, including games and surgical simulations.

Recent single image super-resolution research plays a crucial role in fields such as medical imaging, satellite imaging, and surveillance. Although the VDSR(Very Deep Super Resolution) model achieves excellent restoration performance through deep convolutional neural networks(CNN) and residual learning , its reliance on pixel-level MSE loss limits its ability to recover high-frequency details—resulting in blurry images, especially at 3× and 4× upscaling. To overcome this limitation, we propose a method that integrates an edge loss term into the VDSR framework. By using a Sobel filter, we extract horizontal and vertical edge maps and compute the L1 loss between the edge maps of the super-resolved image and the ground truth. This additional loss encourages the model to restore both pixel accuracy and fine structural details. Experimental results on the Set5 dataset under multiple upscaling scales indicate that performance gains in PSNR and SSIM become more pronounced as the scaling factor increases.

In Computer Vision, Super-Resolution is a key technique used to convert low-resolution images into high- resolution images. Traditional bicubic interpolation-based methods suffer from detail loss and exhibit limitations in restoring edges and textures. In this paper, we propose a method that optimizes upscaling filters using a learnable transposed convolution and applies subsequent convolutional layers to restore fine details.The proposed method performs direct upscaling using a single transposed convolution layer, while subsequent convolutional layers are designed to minimize checkerboard artifacts and correct pixel imbalance. Additionally, Super-Resolution Convolutional Neural Network is applied in the final stage to enhance fine structures and further improve image quality. Experimental results demonstrate that our model outperforms conventional bicubic interpolation and standard transposed convolution methods in terms of Peak Signal-to-Noise Ratio and Structural Similarity Index Measure, producing visually sharper results. This study is expected to provide an effective solution for Super-Resolution image reconstruction and various Computer Vision applications.