检索结果-内蒙古大学图书馆

Backdoor learning curves: explaining backdoor poisoning beyond influence functions

INTERNATIONAL JOURNAL OF MACHINE LEARNING AND CYBERNETICS 2025年第3期16卷 1779-1804页

作者： Cina, Antonio Emanuele Grosse, Kathrin Vascon, Sebastiano Demontis, Ambra Biggio, Battista Roli, Fabio Pelillo, Marcello Univ Genoa DIBRIS Genoa Italy CaFoscari Univ Venice DAIS Venice Italy Ecole Polytech Fedeerale Lausanne VITA Lab Lausanne Switzerland Univ Cagliari DIEE Cagliari Italy

Backdoor attacks inject poisoning samples during training, with the goal of forcing a machine learning model to output an attacker-chosen class when presented with a specific trigger at test time. Although backdoor attacks have been demonstrated in a variety of settings and against different models, the factors affecting their effectiveness are still not well understood. In this work, we provide a unifying framework to study the process of backdoor learning under the lens of incremental learning and influence functions. We show that the effectiveness of backdoor attacks depends on (i) the complexity of the learning algorithm, controlled by its hyperparameters;(ii) the fraction of backdoor samples injected into the training set;and (iii) the size and visibility of the backdoor trigger. These factors affect how fast a model learns to correlate the presence of the backdoor trigger with the target class. Our analysis unveils the intriguing existence of a region in the hyperparameter space in which the accuracy of clean test samples is still high while backdoor attacks are ineffective, thereby suggesting novel criteria to improve existing defenses.

关键词： Backdoor poisoning influence functions Poisoning Machine learning Adversarial machine learning Security

来源：评论

学校读者我要写书评

暂无评论

Regularization of Hole-Drilling Residual Stress Measurements with Eccentric Holes: An Approach with influence functions

引用

JOURNAL OF MATERIALS ENGINEERING AND PERFORMANCE 2024年第15期33卷 7652-7658页

作者： Beghini, M. Bertini, L. Cococcioni, M. Grossi, T. Santus, C. Benincasa, A. Univ Pisa Dipartimento Ingn Civile & Ind Pisa PI Italy Univ Pisa Dipartimento Ingn Informaz Pisa PI Italy SINT Technol Srl Calenzano FI Italy

The hole-drilling method is one of the most widespread techniques to measure residual stresses. Since the introduction of the Integral Method to evaluate non-uniform stress distributions, there has been a considerable improvement in the instrumentation technology, as step increments of about 10 microns are now achievable. However, that spatial resolution makes the ill-posedness of the problem stand out among other sources of uncertainty. As the solution becomes totally dominated by noise, an additional regularization of the problem is needed to obtain meaningful results. Tikhonov regularization is the most common option, as it is also prescribed by the hole-drilling ASTM E837 standard, but it has only been studied in the reference case of a hole with no eccentricity with respect to the strain rosette. A recent work by Schajer addresses the eccentricity problem by defining a correction strategy that transforms strain measurements, allowing one to obtain the solution with the usual decoupled equations. In this work, Tikhonov regularization is applied to the eccentric hole case through the influence functions approach, in order to avoid the introduction of new error-compensating functions and bias-prone interpolations. Some useful general considerations for a practical implementation of the procedure and an experimental test case on an aluminum specimen are presented.

关键词： eccentricity hole-drilling influence functions regularization residual stress

来源：评论

学校读者我要写书评

暂无评论

Strain concentration factor of heterogeneous materials and analytical influence functions based on Eshelby tensor

引用

Theoretical & Applied Mechanics Letters 2024年第4期14卷 306-313页

作者： Shanqiao Huang Zifeng Yuan HEDPS Center for Applied Physics and TechnologyDepartment of Mechanics and Engineering SciencePeking UniversityBeijing 100871China State Key Laboratory for Turbulence and Complex Systems Peking UniversityBeijing 100871China

In this manuscript,Eshelby tensor is employed to assess the strain concentration that arises in the matrix phase at the interface,offering precise values and locations of maximum strain under specific loading conditions for both spherical and cylindrical *** compared to numerical simulation results,the analytical predictions grounded in the Eshelby tensor exhibit satisfactory *** an analytical calculation method based on Eshelby tensor for the elastic strain influence functions of reduced-order homogenization(ROH)method is developed and adopted on particle-reinforced and fibrous composites,presenting its feasibility and advantage on off-line stage calculation of ROH *** error analyses between analytical and numerical results are *** numerical results also exhibit the necessity of finer interface partitioning to obtain the response on micro-scale with higher resolution.

关键词： Eshelby tensor Strain concentration Composite structures Reduced-order-homogenization influence functions

来源：评论

学校读者我要写书评

暂无评论

Revisiting the fragility of influence functions

引用

NEURAL NETWORKS 2023年第1期162卷 581-588页

作者： Epifano, Jacob R. Ramachandran, Ravi P. Masino, Aaron J. Rasool, Ghulam Rowan Univ Dept Elect & Comp Engn 201 Mullica Hill Rd Glassboro NJ 08028 USA Univ Penn Dept Biostat Perelman Sch Med InformatEpidemiol 423 Guardian Dr Philadelphia PA 19104 USA H Lee Moffitt Canc Ctr & Res Inst Dept Machine Learning 12902 USF Magnolia Dr Tampa FL 33612 USA

In the last few years, many works have tried to explain the predictions of deep learning models. Few methods, however, have been proposed to verify the accuracy or faithfulness of these explanations. Recently, influence functions, which is a method that approximates the effect that leave-one-out training has on the loss function, has been shown to be fragile. The proposed reason for their fragility remains unclear. Although previous work suggests the use of regularization to increase robustness, this does not hold in all cases. In this work, we seek to investigate the experiments performed in the prior work in an effort to understand the underlying mechanisms of influence function fragility. First, we verify influence functions using procedures from the literature under conditions where the convexity assumptions of influence functions are met. Then, we relax these assumptions and study the effects of non-convexity by using deeper models and more complex datasets. Here, we analyze the key metrics and procedures that are used to validate influence functions. Our results indicate that the validation procedures may cause the observed fragility.

关键词： Machine learning Supervised learning Deep learning Explainable AI influence functions Bayesian neural networks

来源：评论

学校读者我要写书评

暂无评论

Evaluating the Impact of Local Differential Privacy on Utility Loss via influence functions

Evaluating the Impact of Local Differential Privacy on Utili...

引用

International Joint Conference on Neural Networks (IJCNN)

作者： Carey, Alycia N. Van, Minh-Hao Wu, Xintao Univ Arkansas Elect Engn & Comp Sci Fayetteville AR 72701 USA

ISBN: (纸本)9798350359329;9798350359312

How to properly set the privacy parameter in differential privacy has been an open question in DP research since it was first proposed in 2006. In this work, we demonstrate the ability of influence functions to offer insight into how a specific privacy parameter value will affect a model's test loss in the randomized response-based local DP setting. Our proposed method allows a data curator to select the privacy parameter best aligned with their allowed privacy-utility trade-off without requiring heavy computation such as extensive model retraining and data privatization. We consider multiple common randomization scenarios, such as performing randomized response over the features, and/or over the labels, as well as the more complex case of applying a class-dependent label noise correction method to offset the noise incurred by randomization. Through empirical evaluations we show that for both binary and multi-class settings, influence functions are able to approximate the true change in test loss that occurs when randomized response is applied over features and/or labels with small mean absolute error,especially in cases where noise correction methods are applied.

关键词： differential privacy randomized response influence functions

来源：评论

学校读者我要写书评

暂无评论

influence functions for a 3D full-space under bilinear stationary loads

引用

ENGINEERING ANALYSIS WITH BOUNDARY ELEMENTS 2021年 130卷 286-299页

作者： Romanini, E. Labaki, J. Vasconcelos, A. C. A. Mesquita, E. Fed Univ South Mato Grosso Tres Lagoas MS Brazil Univ Estadual Campinas Sch Mech Engn Campinas Brazil

This manuscript brings the derivation of influence functions for a three-dimensional full-space under bilinearlydistributed time-harmonic loads. The differential equations describing the medium are decomposed in terms of uncoupled vector fields. A double Fourier transform allows the system of equations to be solved algebraically in the transformed space, where the bilinear-loading boundary conditions are imposed. The resulting displacement and stress solutions are presented in terms of double improper integrals to be evaluated numerically. The manuscript brings selected results from the evaluation of these solutions. These influence functions can be used in boundary element models of elastodynamic problems to yield computationally-efficient solutions and improved representation of sharply-varying contact traction fields.

关键词： influence functions Green's functions boundary element method Soil-structure interaction

来源：评论

学校读者我要写书评

暂无评论

"How Biased are Your Features?": Computing Fairness influence functions with Global Sensitivity Analysis 23

"How Biased are Your Features?": Computing Fairness Influenc...

引用

6th ACM Conference on Fairness, Accountability, and Transparency (FAccT)

作者： Ghosh, Bishwamittra Basu, Debabrota Meel, Kuldeep S. Natl Univ Singapore Singapore Singapore Univ Lille Equipe Scool INRIA UMR 9189 CRIStALCNRS Lille France

ISBN: (纸本)9781450372527

Fairness in machine learning has attained significant focus due to the widespread application in high-stake decision-making tasks. Unregulated machine learning classifiers can exhibit bias towards certain demographic groups in data, thus the quantification and mitigation of classifier bias is a central concern in fairness in machine learning. In this paper, we aim to quantify the influence of different features in a dataset on the bias of a classifier. To do this, we introduce the Fairness influence Function (FIF). This function breaks down bias into its components among individual features and the intersection of multiple features. The key idea is to represent existing group fairness metrics as the difference of the scaled conditional variances in the classifier's prediction and apply a decomposition of variance according to global sensitivity analysis. To estimate FIFs, we instantiate an algorithm FairXplainer that applies variance decomposition of classifier's prediction following local regression. Experiments demonstrate that FairXplainer captures FIFs of individual feature and intersectional features, provides a better approximation of bias based on FIFs, demonstrates higher correlation of FIFs with fairness interventions, and detects changes in bias due to fairness affirmative/punitive actions in the classifier. The code is available at https://***/ReAILe/bias- explainer. The extended version of the paper is at https://***/pdf/2206. ***.

关键词： Fair Machine Learning Bias Explainability Global Sensitivity Analysis Variance Decomposition influence functions

来源：评论

学校读者我要写书评

暂无评论

Efficient Approximate Predictive Inference Under Feedback Covariate Shift with influence functions 12

Efficient Approximate Predictive Inference Under Feedback Co...

引用

12th Symposium on Conformal and Probabilistic Prediction with Applications (COPA)

作者： Prinster, Drew Saria, Suchi Liu, Anqi Johns Hopkins Univ Dept Comp Sci Baltimore MD 21218 USA

We propose JAWA-FCS, which uses higher-order influence functions to approximate predictive intervals of the (previous) jackknife+ weighted for feedback covariate shift for further computational efficiency (no retraini... 详细信息

关键词： Jackknife feedback covariate shift influence functions conformal prediction biomolecular design

来源：评论

学校读者我要写书评

暂无评论

influence functions FOR DIMENSION REDUCTION METHODS

引用

BULLETIN OF THE AUSTRALIAN MATHEMATICAL SOCIETY 2021年第3期103卷 517-519页

作者： Smith, Jodie Ann La Trobe Univ Sch Engn & Math Sci Melbourne Vic 3086 Australia

An abstract is not available for this content so a preview has been provided. As you have access to this article, a PDF of this content is available in through the ‘Save PDF’ action button. <div

An abstract is not available for this content so a preview has been provided. As you have access to this article, a PDF of this content is available in through the ‘Save PDF’ action button.

关键词： dimension reduction eigen-subspace influence analysis influence functions influential observations linear discriminant analysis principal Hessian directions sliced inverse regression

来源：评论

学校读者我要写书评

暂无评论

Using influence functions to identify potential improvements for synthetic data generation 4

Using influence functions to identify potential improvements...

引用

Conference on Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications IV

作者： Glandon, S. Ross Agrawal, Rajeev Cheng, Jing-Ru C. Maxwell, Andrew US Army Engineer Res & Dev Ctr Vicksburg MS 39180 USA

ISBN: (纸本)9781510651036;9781510651029

Computer vision, enabled by artificial intelligence and deep learning, has a nearly limitless number of possible applications, military and civilian. Object detection methods are a particularly notable type of computer vision, with broad usefulness in a variety of systems, such as autonomous vehicles, robotics, and security. Development of effective object detection methods faces many challenges;one such challenge of significance is a lack of good labeled data for the target domain, as hand labeled data is time-consuming and expensive to produce. Synthetic data generation seeks to solve this problem by programmatically generating both training data and labels simultaneously, allowing for the creation of arbitrarily large training datasets. However, synthetic data has several drawbacks - generating realistic imagery is challenging and computationally expensive, and models trained with synthetic data frequently suffer in accuracy when applied to real test data. In our research, we use model explainability techniques to connect model predictions back to the model training data, in order to identify the most important features that need to be represented accurately in synthetic training data. influence functions score model training samples based on how influential each sample was to a particular prediction, by approximating the effect of retraining the model after leaving each individual training sample out of the training set. In this work, we seek to extend influence functions to identify the most valuable features in real and synthetic training data for use in improving our synthetic data generation tools.

关键词： influence functions synthetic data machine learning computer vision

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：