{"title":"提高对抗鲁棒性的瓦瑟斯坦切片对抗训练","authors":"Woojin Lee, Sungyoon Lee, Hoki Kim, Jaewook Lee","doi":"10.1007/s12652-024-04791-1","DOIUrl":null,"url":null,"abstract":"<p>Recently, deep-learning-based models have achieved impressive performance on tasks that were previously considered to be extremely challenging. However, recent works have shown that various deep learning models are susceptible to adversarial data samples. In this paper, we propose the sliced Wasserstein adversarial training method to encourage the logit distributions of clean and adversarial data to be similar to each other. We capture the dissimilarity between two distributions using the Wasserstein metric and then align distributions using an end-to-end training process. We present the theoretical background of the motivation for our study by providing generalization error bounds for adversarial data samples. We performed experiments on three standard datasets and the results demonstrate that our method is more robust against white box attacks compared to previous methods.</p>","PeriodicalId":14959,"journal":{"name":"Journal of Ambient Intelligence and Humanized Computing","volume":null,"pages":null},"PeriodicalIF":0.0000,"publicationDate":"2024-05-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Sliced Wasserstein adversarial training for improving adversarial robustness\",\"authors\":\"Woojin Lee, Sungyoon Lee, Hoki Kim, Jaewook Lee\",\"doi\":\"10.1007/s12652-024-04791-1\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<p>Recently, deep-learning-based models have achieved impressive performance on tasks that were previously considered to be extremely challenging. However, recent works have shown that various deep learning models are susceptible to adversarial data samples. In this paper, we propose the sliced Wasserstein adversarial training method to encourage the logit distributions of clean and adversarial data to be similar to each other. We capture the dissimilarity between two distributions using the Wasserstein metric and then align distributions using an end-to-end training process. We present the theoretical background of the motivation for our study by providing generalization error bounds for adversarial data samples. We performed experiments on three standard datasets and the results demonstrate that our method is more robust against white box attacks compared to previous methods.</p>\",\"PeriodicalId\":14959,\"journal\":{\"name\":\"Journal of Ambient Intelligence and Humanized Computing\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2024-05-05\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Journal of Ambient Intelligence and Humanized Computing\",\"FirstCategoryId\":\"94\",\"ListUrlMain\":\"https://doi.org/10.1007/s12652-024-04791-1\",\"RegionNum\":3,\"RegionCategory\":\"计算机科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"Computer Science\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of Ambient Intelligence and Humanized Computing","FirstCategoryId":"94","ListUrlMain":"https://doi.org/10.1007/s12652-024-04791-1","RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"Computer Science","Score":null,"Total":0}
Sliced Wasserstein adversarial training for improving adversarial robustness
Recently, deep-learning-based models have achieved impressive performance on tasks that were previously considered to be extremely challenging. However, recent works have shown that various deep learning models are susceptible to adversarial data samples. In this paper, we propose the sliced Wasserstein adversarial training method to encourage the logit distributions of clean and adversarial data to be similar to each other. We capture the dissimilarity between two distributions using the Wasserstein metric and then align distributions using an end-to-end training process. We present the theoretical background of the motivation for our study by providing generalization error bounds for adversarial data samples. We performed experiments on three standard datasets and the results demonstrate that our method is more robust against white box attacks compared to previous methods.
期刊介绍:
The purpose of JAIHC is to provide a high profile, leading edge forum for academics, industrial professionals, educators and policy makers involved in the field to contribute, to disseminate the most innovative researches and developments of all aspects of ambient intelligence and humanized computing, such as intelligent/smart objects, environments/spaces, and systems. The journal discusses various technical, safety, personal, social, physical, political, artistic and economic issues. The research topics covered by the journal are (but not limited to):
Pervasive/Ubiquitous Computing and Applications
Cognitive wireless sensor network
Embedded Systems and Software
Mobile Computing and Wireless Communications
Next Generation Multimedia Systems
Security, Privacy and Trust
Service and Semantic Computing
Advanced Networking Architectures
Dependable, Reliable and Autonomic Computing
Embedded Smart Agents
Context awareness, social sensing and inference
Multi modal interaction design
Ergonomics and product prototyping
Intelligent and self-organizing transportation networks & services
Healthcare Systems
Virtual Humans & Virtual Worlds
Wearables sensors and actuators