A Boundary-Zone Guided Diffusion Model for Adversarial Training with Label Uncertainty in Mobile Networks

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Communication Engineering > Electronic Thesis & Dissertation > Item 987654321/98134

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98134

Title:	A Boundary-Zone Guided Diffusion Model for Adversarial Training with Label Uncertainty in Mobile Networks
Authors:	馬寧;Ma, Ning
Contributors:	通訊工程學系
Keywords:	生成式 AI;表格型去除雜訊擴散機率模型(TabDDPM）;決策邊界;對抗訓練;標籤不確定性;Generative AI;tabular denoising diffusion probabilistic model (TabDDPM);decision boundary;adversarial training;label uncertainty
Date:	2025-08-16
Issue Date:	2025-10-17 12:23:54 (UTC+8)
Publisher:	國立中央大學
Abstract:	下一代行動網路在關鍵任務上高度依賴機器學習，因此模型對於對抗性攻擊的穩健性至關重要。我們提出了 AdvTabDDPM，一個基於擴散模型的框架，用於混合型表格資料中生成邊界敏感的潛在對抗性樣本。透過定義動態邊界區域並在對抗訓練中引入標籤不確定性（label uncertainty），我們的方法能精確針對低信心區域，促使決策邊界更加平滑且可靠。實驗結果顯示，AdvTabDDPM 能將生成的樣本集中於決策邊界附近，並將整體準確率從 82.06% 提升至 92.57%，在所有評估方法中最接近未受擾動資料的準確率 93.87%，並優於 FGSM、PGD 與 CW。這些結果證明，結合邊界導向的樣本生成與標籤不確定性，能為網路機器學習模型提供有效的穩健性與準確性提升策略。;Next-generation mobile networks rely on machine learning for critical tasks, making model robustness against adversarial attacks essential. We propose AdvTabDDPM, a diffusion-based framework that generates boundary-sensitive potential adversarial examples in mixed-type tabular datasets. By defining a dynamic boundary zone and incorporating label uncertainty during adversarial training, our method precisely targets low-confidence regions, encouraging smoother and more reliable decision boundaries. Experiments show that AdvTabDDPM concentrates examples near decision boundaries and improves overall accuracy from 82.06% to 92.57%, the closest among evaluated methods to the clean data accuracy of 93.87%, outperforming FGSM, PGD, and CW. These results demonstrate that combining boundary-guided generation with label uncertainty provides an effective strategy for robust and accurate network ML models.
Appears in Collections:	[Graduate Institute of Communication Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	8	View/Open

社群 sharing

Loading...