logo
科言猫学术研究的AI总结
language
手机端手机端
  • Phone Feature微信小程序
简体中文简体中文
notice消息通知
博客博客
delete
arrow
返回

Variance-Based Defense Against Blended Backdoor Attacks

delete2026-01-01
delete0
PRE
AI
S
Sujeevan Aseervatham *
A
Achraf Kerzazi
Y
Younès Bennani
DOI：10.1007/978-3-032-06096-9_13delete
delete原文链接
delete原文求助
delete分享
delete收藏
摘要

摘要

En 中文
后门攻击代表了一种针对AI模型的隐蔽且有效的网络攻击类别，主要因其隐蔽性。模型在干净数据上表现正常，但仅在攻击者将特定触发器嵌入输入时才表现出恶意行为。这种攻击在训练阶段执行，攻击者通过嵌入模式并修改标签至选定目标来污染训练数据的一个小子集。目标是使模型将模式与目标标签关联，同时保持未修改数据上的正常性能。已提出多种防御机制来净化训练数据集。然而，这些方法通常依赖干净数据集来检测统计异常，这在现实场景中可能不可行，因为数据集可能缺失或被篡改。为解决此限制，我们提出一种新颖的防御方法：在给定数据集上训练模型，检测被污染的类别，并在识别被污染实例之前提取攻击触发器的关键部分。该方法通过明确揭示触发器的有害部分增强了可解释性。我们的方法的有效性通过在知名图像数据集上的实验评估以及与三种最新算法（SCAn、ABL和AGPD）的比较分析得到验证。
Keyword:
Machine Learning
Data Poisoning
Backdoor Mitigation

期刊

M
MACHINE LEARNING AND KNOWLEDGE DISCOVERY IN DATABASES. RESEARCH TRACK, ECML PKDD 2025, PT V
IF:
0
论文数:
26
被引数:
0

机构

C
centre national de la recherche scientifique (cnrs)
学者数:
24.5W
论文数: 18.2W
被引数: 279
C
cnrs - institute for information sciences & technologies (ins2i)
学者数:
590
论文数: 443
被引数: 0
引用论文

引用论文

SentiNet: Detecting Localized Universal Attacks Against Deep Learning Systems
err2020 IEEE Security and Privacy Workshops (SPW)
IF0
err2020-05-01
err0
errOAAI
errEdward Chou; Florian Tramer; Giancarlo Pellegrino
err分享
err收藏
STRIP
errProceedings of the 35th Annual Computer Security Applications Conference
IF0
err2019-12-09
err0
PREAI
errYansong Gao; Change Xu; Derui Wang; Shiping Chen; Damith C. Ranasinghe; Surya Nepal
err分享
err收藏
A taxonomy and survey of attacks against machine learning
errCOMPUTER SCIENCE REVIEW
IF12.7
err2019-11-01
err167
errOAAI
errPitropakis, Nikolaos; Panaousis, Emmanouil; Giannetsos, Thanassis; Anastasiadis, Eleftherios; Loukas, George
err分享
err收藏
A Threshold Selection Method from Gray-Level Histograms一种基于灰度直方图的阈值选择方法
errIEEE Transactions on Systems, Man, and Cybernetics
IF0
err1979-01-01
err0
PREAI
errNobuyuki Otsu
err分享
err收藏
The "Beatrix" Resurrections: Robust Backdoor Detection via Gram Matrices
errProceedings 2023 Network and Distributed System Security Symposium
IF0
err2023-01-01
err0
errOAAI
errWanlun Ma; Derui Wang; Ruoxi Sun; Minhui Xue; Sheng Wen; Yang Xiang
err分享
err收藏
Dynamic Backdoor Attacks Against Machine Learning Models
err2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P)
IF0
err2022-06-01
err0
errOAAI
errAhmed Salem; Rui Wen; Michael Backes; Shiqing Ma; Yang Zhang
err分享
err收藏
Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based LocalizationGrad-cam: 通过基于梯度的定位从深度网络进行视觉解释
errINTERNATIONAL JOURNAL OF COMPUTER VISION
IF9.3
err2019-10-11
err8.0K
errOAAI
errSelvaraju, Ramprasaath R.; Cogswell, Michael; Das, Abhishek; Vedantam, Ramakrishna; Parikh, Devi; Batra, Dhruv
err分享
err收藏
Rethinking the Backdoor Attacks’ Triggers: A Frequency Perspective
err2021 IEEE/CVF International Conference on Computer Vision (ICCV)
IF0
err2021-10-01
err0
errOAAI
errYi Zeng; Won Park; Z. Morley Mao; Ruoxi Jia
err分享
err收藏
A Survey on Security Threats and Defensive Techniques of Machine Learning: A Data Driven View机器学习的安全威胁和防御技术综述: 数据驱动的观点
errIEEE ACCESS
IF3.6
err2018-01-01
err263
errOAAI
errLiu, Qiang; Li, Pan; Zhao, Wentao; Cai, Wei; Yu, Shui; Leung, Victor C. M.
err分享
err收藏
Identity Mappings in Deep Residual Networks
errLecture Notes in Computer Science
IF0
err2016-09-17
err0
PREAI
errKaiming He; Xiangyu Zhang; Shaoqing Ren; Jian Sun
err分享
err收藏
没有更多内容