2026, Vol.31, No.8

Journal of The Korea Society of Computer and Information 2025 KCI Impact Factor : 1.01

pISSN : 1598-849X / eISSN : 2383-9945
https://journal.kci.go.kr/jksci

Multi-cloud environments improve resource utilization but also increase management complexity and security risks caused by configuration errors. This study proposes a security-enhanced meta-cloud architecture that integrates visual infrastructure design, Terraform-based Infrastructure as Code (IaC) generation, pre- and post-deployment security validation, and multi-cloud operations into a unified framework. The proposed architecture applies consistent security policies across AWS, Microsoft Azure, and Naver Cloud Platform (NCP), demonstrating consistent operation in heterogeneous cloud environments. Experiments involving 144 non-expert participants achieved an average architecture implementation completeness of 66.8%, while the automated security inspection process significantly improved operational efficiency compared with conventional manual approaches. The results demonstrate that the proposed architecture can effectively enhance both the security posture and operational efficiency of multi-cloud environments.

In this paper, we propose a class-balanced evaluation protocol across data scales based on the VTAB-1k benchmark to analyze Parameter-Efficient Fine-Tuning (PEFT) under the extreme data scarcity and class imbalance of real-world domains such as medical and satellite imagery. The pipeline independently controls data scale (10%–100%) and class balance, benchmarking Adapter, LoRA, and Visual Prompt Tuning (VPT) on 19 tasks over a frozen ViT backbone, together with full fine-tuning and linear-probing baselines and a recent variant, DoRA. A clear performance crossover is observed: VPT, strong in specialized and structured domains, dominates below the 50% scale, whereas LoRA, strong in natural domains, prevails as data increases. Class balancing trades a small amount of overall Top-1 accuracy for higher balanced accuracy in low-data regimes, and its effect vanishes at larger scales as the balanced subset converges to the original distribution. These controlled results demonstrate that no single PEFT method is universally optimal, and the resulting guidelines support the practical deployment of PEFT in data-limited applications such as medical imaging, remote sensing, and industrial inspection.

In this paper, we propose SADA-DFQ, a generative data-free quantization (DFQ) framework that restores a low-bit network without any access to the original training data, as required in privacy-sensitive domains such as medical imaging and security. Although generator-based DFQ offers high sample diversity, its accuracy still lags behind. We attribute this gap to two factors: an unreliable intermediate-layer transfer signal at low bit-width, where rounding and clipping errors prevent the quantized model from reproducing the full-precision (FP) model's activation magnitudes, and the limited informativeness of the synthetic dataset used for fine-tuning. To address the first, Self-Attention Correlation Distillation (SACD) reformulates each intermediate feature into multi-head self-attention and transfers the relational structure among channel tokens, which tends to be less sensitive to quantization error and thus provides a more stable signal. To address the second, Difficulty-Aware Sample Generation (DASG) re-weights the synthetic dataset by sample difficulty to emphasize examples near the decision boundary. On CIFAR-100, SADA-DFQ attains the best accuracy among the compared generator-based DFQ methods at both 3-bit and 4-bit, improving Top-1 accuracy by about 2.6 percentage points over AdaDFQ at 3-bit.