Convolution Based Unlearnable
Convolution-based unlearnable datasets (UDs) aim to create training data that severely hinders the performance of deep learning models, primarily to address data privacy concerns in publicly available datasets. Current research focuses on developing robust UD generation techniques using class-wise convolutions with randomly generated filters, and exploring defenses against these UDs, such as pixel-based image transformations designed to mitigate the impact of multiplicative noise. The effectiveness of these UDs and defenses is evaluated across various datasets and architectures, including ResNets, VGGs, and EfficientNets, highlighting the need for robust data protection strategies in machine learning.
Papers
November 4, 2024
November 30, 2023