Image Perturbation

Image perturbation involves subtly altering images to understand how these changes affect model performance, explain model predictions, or create adversarial attacks. Current research focuses on developing robust perturbation methods for various tasks, including explanation generation, adversarial attack design, and data protection, often employing techniques like attention mechanisms, generative models (e.g., diffusion models, GANs), and variational autoencoders. These advancements are crucial for improving model explainability, enhancing model security against malicious attacks, and developing more reliable and robust computer vision systems across diverse applications.

Papers

November 28, 2021

Detecting Adversaries, yet Faltering to Noise? Leveraging Conditional Variational AutoEncoders for Adversary Detection in the Presence of Noisy Images
Dvij Kalaria, Aritra Hazra, Partha Pratim Chakrabarti
Deep Neural Network Adversarial Attack Deep Learning Model Industrial Disturbing Noise Noisy Image Conditional Variational Image Perturbation Adversary Model

November 7, 2021

Look at the Variance! Efficient Black-box Explanations with Sobol-based Sensitivity Analysis
Thomas Fel, Remi Cadene, Mathieu Chalvidal, Matthieu Cord, David Vigouroux, Thomas Serre
Line by Line Explanation High Dimension Variance Information Sensitivity Analysis High Dimensional Problem Image Perturbation

Image Perturbation

Papers

Detecting Adversaries, yet Faltering to Noise? Leveraging Conditional Variational AutoEncoders for Adversary Detection in the Presence of Noisy Images

Look at the Variance! Efficient Black-box Explanations with Sobol-based Sensitivity Analysis