Imperceptible Perturbation

Imperceptible perturbations are subtle alterations to input data designed to mislead machine learning models without noticeably changing the input itself. Current research focuses on understanding and mitigating the vulnerability of various models, including convolutional neural networks (CNNs), graph neural networks (GNNs), and diffusion models, to these attacks across different data types (images, tabular data). This research is crucial for improving the robustness and reliability of machine learning systems in security-sensitive applications and for developing more trustworthy AI. A key challenge lies in defining and measuring "imperceptibility" across diverse data modalities, and in balancing the trade-off between attack effectiveness and the undetectability of the perturbation.