Active subspace of neural networks: structural analysis and universal attacks

DOI10.1137/19M1296070zbMATH Open1486.90152arXiv1910.13025OpenAlexW3095125464MaRDI QIDQ5037556FDOQ5037556

Authors: Chunfeng Cui, Kaiqi Zhang, T. Daulbaev, Julia Gusak, Ivan Oseledets, Zheng Zhang 

Publication date: 1 March 2022

Published in: SIAM Journal on Mathematics of Data Science (Search for Journal in Brave)

Abstract: Active subspace is a model reduction method widely used in the uncertainty quantification community. In this paper, we propose analyzing the internal structure and vulnerability and deep neural networks using active subspace. Firstly, we employ the active subspace to measure the number of "active neurons" at each intermediate layer and reduce the number of neurons from several thousands to several dozens. This motivates us to change the network structure and to develop a new and more compact network, referred to as {ASNet}, that has significantly fewer model parameters. Secondly, we propose analyzing the vulnerability of a neural network using active subspace and finding an additive universal adversarial attack vector that can misclassify a dataset with a high probability. Our experiments on CIFAR-10 show that ASNet can achieve 23.98$imes$ parameter and 7.30$imes$ flops reduction. The universal active subspace attack vector can achieve around 20% higher attack ratio compared with the existing approach in all of our numerical experiments. The PyTorch codes for this paper are available online.

Full work available at URL: https://arxiv.org/abs/1910.13025

Recommendations

• Gaussian processes with built-in dimensionality reduction: applications to high-dimensional uncertainty propagation
• Gradient-free construction of active subspaces for dimension reduction in complex models with applications to neutronics
• Generalized bounds for active subspaces
• Active subspace methods in theory and practice: applications to kriging surfaces
• Deep UQ: learning deep neural network surrogate models for high dimensional uncertainty quantification

zbMATH Keywords

deep neural networknetwork reductionactive subspaceuniversal adversarial perturbation

Mathematics Subject Classification ID

Nonparametric robustness (62G35) Eigenvalues, singular values, and eigenvectors (15A18) Nonconvex programming, global optimization (90C26)

Cites Work

• GXNOR-Net: training deep neural networks with ternary weights and activations without full-precision memory under a unified discretization framework
• Nonmonotone Spectral Projected Gradient Methods on Convex Sets
• Finding structure with randomness: probabilistic algorithms for constructing approximate matrix decompositions
• Minimization of functions having Lipschitz continuous first partial derivatives
• The Wiener--Askey Polynomial Chaos for Stochastic Differential Equations
• Title not available (Why is that?)
• Real analysis
• Modeling uncertainty in steady state diffusion problems via generalized polynomial chaos
• Applied logistic regression
• Accelerated proximal stochastic dual coordinate ascent for regularized loss minimization
• Active subspaces. Emerging ideas for dimension reduction in parameter studies
• Exploiting active subspaces to quantify uncertainty in the numerical simulation of the hyshot II scramjet
• Erratum: Active Subspace Methods in Theory and Practice: Applications to Kriging Surfaces
• FitNets
• Title not available (Why is that?)
• Frequent directions: simple and deterministic matrix sketching

Cited In (5)

• Model-independent detection of new physics signals using interpretable semisupervised classifier tests
• A Supervised Learning Approach Involving Active Subspaces for an Efficient Genetic Algorithm in High-Dimensional Optimization Problems
• Mappings, dimensionality and reversing out of deep neural networks
• Kernel‐based active subspaces with application to computational fluid dynamics parametric problems using the discontinuous Galerkin method
• On the Deep Active-Subspace Method

Uses Software

• aplore3
• DeepFool
• ImageNet
• Adam
• GXNOR-Net
• BinaryNet
• AMC
• AlexNet
• FitNets
• ProxylessNAS

This page was built for publication: Active subspace of neural networks: structural analysis and universal attacks

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q5037556)