Unlearnable Datasets

Unlearnable datasets aim to protect sensitive data used in training deep learning models by introducing carefully designed perturbations that hinder unauthorized model training while preserving data usability for authorized users. Current research focuses on developing robust perturbation techniques, including those based on nonlinear transformations, semantic image hiding, and class-wise manipulations of 3D point clouds, and exploring the vulnerabilities of existing methods to countermeasures like adversarial training and data augmentation. This field is crucial for advancing data privacy and security in machine learning, with implications for protecting intellectual property and sensitive information in various applications.