Advancing Beyond Identification: Multi-bit Watermark for Large Language Models
About
We show the viability of tackling misuses of large language models beyond the identification of machine-generated text. While existing zero-bit watermark methods focus on detection only, some malicious misuses demand tracing the adversary user for counteracting them. To address this, we propose Multi-bit Watermark via Position Allocation, embedding traceable multi-bit information during language model generation. Through allocating tokens onto different parts of the messages, we embed longer messages in high corruption settings without added latency. By independently embedding sub-units of messages, the proposed method outperforms the existing works in terms of robustness and latency. Leveraging the benefits of zero-bit watermarking, our method enables robust extraction of the watermark without any model access, embedding and extraction of long messages ($\geq$ 32-bit) without finetuning, and maintaining text quality, while allowing zero-bit detection all at the same time. Code is released here: https://github.com/bangawayoo/mb-lm-watermarking
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Fake News Detection | FAKE NEWS | Accuracy93.31 | 66 | |
| Watermark Detection | mmw story | Accuracy99.61 | 48 | |
| Watermark Detection | fake_news | Accuracy97.94 | 48 | |
| Watermark Detection | book_report | Accuracy98.31 | 48 | |
| Watermark Detection | finance_qa | Accuracy92.22 | 48 | |
| Watermark Detection | longform_qa | Accuracy90.13 | 48 | |
| Watermark Detection | dolly_cw | Accuracy90.81 | 48 | |
| Detection Accuracy | mmw story | Accuracy97.66 | 24 | |
| Detection Accuracy | LongForm QA | Accuracy94.16 | 24 | |
| Detection Accuracy | C4 subset | Accuracy95.19 | 24 |