This academic article introduces PosCUDA, a method to create unlearnable audio datasets. The method uses class-wise convolutions on small patches of audio, with the location of these patches based on a private key for each class. This allows the model to learn the relations between positional blurs and labels, while failing to generalize. The authors claim that PosCUDA can achieve unlearnability while maintaining the quality of the original audio datasets and is robust to different audio feature representations such as MFCC and raw audio, as well as different architectures like transformers and convolutional networks.
Publication date: 5 Jan 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2401.02135
