干旱区生态环境知识资源中心(Arid): The influence of dropout and residual connection against membership inference attacks on transformer model: a neuro generative disease case study

Arid

DOI	10.1007/s11042-023-16126-x
	The influence of dropout and residual connection against membership inference attacks on transformer model: a neuro generative disease case study
	Ben Hamida, Sameh; Ben Hamida, Sana; Snoun, Ahmed; Jemai, Olfa; Jemai, Abderrazek
通讯作者	Ben Hamida, S
来源期刊	MULTIMEDIA TOOLS AND APPLICATIONS
ISSN	1380-7501
EISSN	1573-7721
出版年	2024
卷号	83 期号:6 页码:16231-16253
英文摘要	Alzheimer's patients necessitate consistent support from caregivers or family members, highlighting the urgency for advanced technologies to aid in their daily lives through early disease detection. Consequently, there has been substantial research and development of machine learning-based systems aimed at assisting Alzheimer's patients. However, ensuring the protection of the sensitive and personal data utilized in these systems remains a critical concern. In this context, Membership Inference Attack poses a severe threat to the privacy of targeted models. This research focuses on enhancing the preservation of data privacy during the training phase. We conducted vulnerability testing on a Transformer deep-learning model against Membership Inference Attack and developed a defense strategy to mitigate its impact. To achieve this objective, we evaluated the studied attack on Transformer model using two datasets: DemCare and Oasis. These datasets contain sensitive and personal information, underscoring the need for their utmost protection. Subsequently, we proposed a defense strategy based on dropout and residual connections. Through comparative experiments, our proposed strategy demonstrated significant improvements (i.e. 20.97% and 18.43%) over the previous model, providing efficient results. Thus, we can confidently conclude that our defense approach enhances data privacy and effectively mitigates the impact of the analyzed attack.
英文关键词	ML Security and privacy Transformer MIA Alzheimer Dropout Residual connection
类型	Article
语种	英语
收录类别	SCI-E
WOS记录号	WOS:001027492600010
WOS类目	Computer Science, Information Systems ; Computer Science, Software Engineering ; Computer Science, Theory & Methods ; Engineering, Electrical & Electronic
WOS研究方向	Computer Science ; Engineering
资源类型	期刊论文
条目标识符	http://119.78.100.177/qdio/handle/2XILL650/404942
推荐引用方式 GB/T 7714	Ben Hamida, Sameh,Ben Hamida, Sana,Snoun, Ahmed,et al. The influence of dropout and residual connection against membership inference attacks on transformer model: a neuro generative disease case study[J],2024,83(6):16231-16253.
APA	Ben Hamida, Sameh,Ben Hamida, Sana,Snoun, Ahmed,Jemai, Olfa,&Jemai, Abderrazek.(2024).The influence of dropout and residual connection against membership inference attacks on transformer model: a neuro generative disease case study.MULTIMEDIA TOOLS AND APPLICATIONS,83(6),16231-16253.
MLA	Ben Hamida, Sameh,et al."The influence of dropout and residual connection against membership inference attacks on transformer model: a neuro generative disease case study".MULTIMEDIA TOOLS AND APPLICATIONS 83.6(2024):16231-16253.