Protecting Language Generation Models via Invisible Watermarking

Authors: Xuandong Zhao, Yu-Xiang Wang, Lei Li

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that GINSEW can effectively identify instances of IP infringement with minimal impact on the generation quality of protected APIs.
Researcher Affiliation Academia 1Department of Computer Science, UC Santa Barbara. Correspondence to: Xuandong Zhao <xuandongzhao@cs.ucsb.edu>, Yu-Xiang Wang <yuxiangw@cs.ucsb.edu>, Lei Li <leili@cs.ucsb.edu>.
Pseudocode Yes Algorithm 1 Watermarking process; Algorithm 2 Watermark detection; Algorithm 3 Watermark detection with text alone.
Open Source Code Yes 1Our source code is available at https://github.com/ Xuandong Zhao/ginseq.
Open Datasets Yes In the machine translation task, we utilize the IWSLT14 and WMT14 datasets (Cettolo et al., 2014; Bojar et al., 2014), specifically focusing on German (De) to English (En) translations. For the story generation task, we use the ROCstories (Mostafazadeh et al., 2016) corpus.
Dataset Splits Yes We adopt the official split of train/valid/test sets. There are 90,000 samples in the train set, and 4081 samples in the validation and test sets.
Hardware Specification Yes All experiments are conducted on an Amazon EC2 P3 instance equipped with four NVIDIA V100 GPUs.
Software Dependencies No The paper mentions that the implementation is based on 'fairseq', but does not provide specific version numbers for software dependencies like fairseq itself, Python, PyTorch, or CUDA.
Experiment Setup Yes We use the Adam optimizer (Kingma & Ba, 2015) with β = (0.9, 0.98) and set the learning rate to 0.0005. Additionally, we incorporate 4,000 warm-up steps. The learning rate then decreases proportionally to the inverse square root of the step number. By default, we use beam search as the decoding method (beam size = 5).