ZeroMark: Towards Dataset Ownership Verification without Disclosing Watermark

Authors: Junfeng Guo, Yiming Li, Ruibo Chen, Yihan Wu, chenxi liu, Heng Huang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on benchmark datasets verify the effectiveness of our Zero Mark and its resistance to potential adaptive attacks.
Researcher Affiliation Academia Junfeng Guo1, , Yiming Li2, , Ruibo Chen1, Yihan Wu1, Chenxi Liu1, Heng Huang1 1Department of Computer Science, Institute of Health Computing University of Maryland College Park 2College of Computing and Data Science, Nanyang Technology University {gjf2023,ruibo,yihanwu,chenxi,heng}@umd.edu; liyiming.tech@gmail.com
Pseudocode Yes Algorithm 1 The main process of our Zero Mark.
Open Source Code Yes The codes for reproducing our main experiments are publicly available at Git Hub.
Open Datasets Yes Currently, there are many (high-quality) public datasets, such as CIFAR [4] and Image Net [5], that can be easily downloaded and used.
Dataset Splits Yes CIFAR-10. CIFAR-10 dataset contains 10 labels, 50,000 training samples, and 10,000 validation samples. The training and validation samples are distributed evenly across each label.
Hardware Specification Yes We use six NVIDIA RTX 2080 Ti GPUs for performing experiments.
Software Dependencies No The paper mentions software like 'Adam optimizer [55]' and 'Backdoor Box2 [54]' but does not provide specific version numbers for general software dependencies (e.g., Python, PyTorch).
Experiment Setup Yes To train DNN models, we use Adam optimizer [55] with the initial learning rate as 0.01.