Manifold Attack

Manifold attacks exploit the geometric structure of data to generate adversarial examples—inputs subtly altered to fool machine learning models while remaining perceptually similar to legitimate data points. Current research focuses on distinguishing between "on-manifold" attacks, which stay within the natural data distribution, and "off-manifold" attacks that deviate from it, employing techniques like projected gradient descent and manifold approximations (e.g., using diffusion maps) to generate and defend against these attacks. Understanding and mitigating the vulnerability of models to these attacks is crucial for improving the robustness and reliability of machine learning systems across various applications, including image classification and natural language processing.

Papers

December 6, 2024

Towards Predicting the Success of Transfer-based Attacks by Quantifying Shared Feature Representations
Ashley S. Dale, Mei Qiu, Foo Bin Che, Thomas Bsaibes, Lauren Christopher, Paul Salama
Joint Representation Attack Success Rate Financial Success Manifold Hypothesis Transfer Based Attack Manifold Attack

March 6, 2024

Effect of Ambient-Intrinsic Dimension Gap on Adversarial Vulnerability
Rajdeep Haldar, Yue Xing, Qifan Song
Adversarial Attack Mixed Effect Adversarial Perturbation Adversarial Vulnerability Intrinsic Dimension Manifold Attack

August 23, 2023

On-Manifold Projected Gradient Descent
Aaron Mahler, Tyrus Berry, Tom Stephens, Harbir Antil, Michael Merritt, Jeanie Schreiber, Ioannis Kevrekidis
Adversarial Attack Adversarial Example Class Manifold Manifold Attack

November 5, 2022

Textual Manifold-based Defense Against Natural Language Adversarial Examples
Dang Minh Nguyen, Luu Anh Tuan
Adversarial Attack Adversarial Example Adversarial Text Adversarial Image Textual Adversarial Example Manifold Attack

December 12, 2021

Interpolated Joint Space Adversarial Training for Robust and Generalizable Defenses
Chun Pong Lau, Jiang Liu, Hossein Souri, Wei-An Lin, Soheil Feizi, Rama Chellappa
Adversarial Attack Adversarial Training Joint Adversarial Manifold Attack

Manifold Attack

Papers

Towards Predicting the Success of Transfer-based Attacks by Quantifying Shared Feature Representations

Effect of Ambient-Intrinsic Dimension Gap on Adversarial Vulnerability

On-Manifold Projected Gradient Descent

Textual Manifold-based Defense Against Natural Language Adversarial Examples

Interpolated Joint Space Adversarial Training for Robust and Generalizable Defenses