SemStamp: A Semantic Watermark with Paraphrastic Robustness for Text Generation

Explore a groundbreaking 15-minute conference talk on SemStamp, a robust sentence-level semantic watermarking algorithm for text generation. Delve into the innovative approach developed by Jack Zhang from the Center for Language & Speech Processing at Johns Hopkins University. Learn how SemStamp addresses the vulnerability of existing watermarking algorithms to paraphrase attacks through its unique design based on locality-sensitive hashing. Discover the algorithm's encoding and hashing process, rejection sampling technique, and margin-based constraint for enhanced robustness. Examine the proposed "bigram" paraphrase attack and its effectiveness against token-level watermarking methods. Gain insights into experimental results demonstrating SemStamp's superior robustness against common and bigram paraphrase attacks, as well as its ability to preserve generation quality compared to previous state-of-the-art methods.