Pith. sign in

REVIEW 15 cited by

Watermarking Text Generated by Black-Box Language Models

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2305.08883 v1 pith:WVWK52DF submitted 2023-05-14 cs.CL cs.AI

classification cs.CLcs.AI
keywords textgeneratedmethodwatermarkblack-boxdetectiondistributionlanguage
verification ladder T0 review T1 audit T2 compute T3 formal

Signed reviews

No signed human review yet.

0 comments
read the original abstract

LLMs now exhibit human-like skills in various fields, leading to worries about misuse. Thus, detecting generated text is crucial. However, passive detection methods are stuck in domain specificity and limited adversarial robustness. To achieve reliable detection, a watermark-based method was proposed for white-box LLMs, allowing them to embed watermarks during text generation. The method involves randomly dividing the model vocabulary to obtain a special list and adjusting the probability distribution to promote the selection of words in the list. A detection algorithm aware of the list can identify the watermarked text. However, this method is not applicable in many real-world scenarios where only black-box language models are available. For instance, third-parties that develop API-based vertical applications cannot watermark text themselves because API providers only supply generated text and withhold probability distributions to shield their commercial interests. To allow third-parties to autonomously inject watermarks into generated text, we develop a watermarking framework for black-box language model usage scenarios. Specifically, we first define a binary encoding function to compute a random binary encoding corresponding to a word. The encodings computed for non-watermarked text conform to a Bernoulli distribution, wherein the probability of a word representing bit-1 being approximately 0.5. To inject a watermark, we alter the distribution by selectively replacing words representing bit-0 with context-based synonyms that represent bit-1. A statistical test is then used to identify the watermark. Experiments demonstrate the effectiveness of our method on both Chinese and English datasets. Furthermore, results under re-translation, polishing, word deletion, and synonym substitution attacks reveal that it is arduous to remove the watermark without compromising the original semantics.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 15 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Selective Disclosure Watermarking for Large Language Models

    cs.CR 2026-07 accept novelty 7.0 of 10

    HeRo recursively partitions the LLM vocabulary into a hierarchy, embedding multi-bit payloads across layers so that verifiers with different keys recover only their authorized portion while preserving the original sam...

  2. Can Watermarking Techniques Help Prevent LLM Model Stealing?

    cs.CR 2026-07 conditional novelty 6.5 of 10

    Softplus-then-perturb with embedding-seeded Gaussian noise defeats PCA/averaging/RPCA dimension-extraction attacks on Mistral-7B and GPT-2 with only modest quality loss.

  3. Observation-Level Watermarking and Detection for Tabular Data

    stat.ME 2026-07 conditional novelty 6.5 of 10

    STAMP embeds Laplace keys via refined empirical CDFs so watermarked tabular rows keep the original law asymptotically and remain detectable from a single observation.

  4. SSTMark: Robust Training-Free Semantic-Level Speech Watermarking

    cs.SD 2026-07 conditional novelty 6.0 of 10

    SSTMark embeds a watermark in AI speech by rewriting its transcript and resynthesizing it, achieving strong average robustness to audio distortions but at the cost of altering the spoken content.

  5. FPEdit: Robust LLM Fingerprinting through Localized Parameter Editing

    cs.CR 2025-08 conditional novelty 6.0 of 10

    FPEdit uses knowledge editing with a promote-suppress objective to embed robust, stealthy natural-language fingerprints into LLMs, achieving 94 to 100 percent retention after fine-tuning while preserving benchmark per...

  6. Optimized Couplings for Watermarking Large Language Models

    cs.CR 2025-05 conditional novelty 6.0 of 10

    For one-shot token watermarking with a min-entropy constraint, the authors derive a closed-form max-min detection rate and show that a correlated channel with balanced random partitions achieves it.

  7. SimMark: A Robust Sentence-Level Similarity-Based Watermarking Algorithm for Large Language Models

    cs.CL 2025-02 conditional novelty 6.0 of 10

    SimMark watermarks LLM text by rejection-sampling sentence pairs so their embedding similarity falls in a secret interval, and detects it with a soft-count z-test that survives paraphrasing.

  8. RAG-WM: An Efficient Black-Box Watermarking Approach for Retrieval-Augmented Generation of Large Language Models

    cs.CR 2025-01 conditional novelty 6.0 of 10

    RAG-WM embeds HMAC-generated entity-relation watermarks into a RAG knowledge base and detects stolen RAGs via black-box queries with a binomial test, showing high success across four LLMs and five datasets.

  9. Invisible Entropy: Towards Safe and Efficient Low-Entropy LLM Watermarking

    cs.CL 2025-05 conditional novelty 5.0 of 10

    A lightweight entropy classifier plus an adaptive threshold method can watermark and detect low-entropy LLM code outputs without querying the original model, matching much larger detectors at 99% fewer detection-phase...

  10. RAFP: Identifying LLM Lineages via Rare-Region Fingerprints

    cs.LG 2025-05 conditional novelty 5.0 of 10

    RoFL (the body's title) shows that rare, low-likelihood prompt-response pairs can serve as stable black-box fingerprints that identify a language model lineage after fine-tuning, LoRA, quantization, and prompt-templat...

  11. CoheMark: A Novel Sentence-Level Watermark for Enhanced Text Quality

    cs.CL 2025-04 conditional novelty 5.0 of 10

    CoheMark embeds a sentence-level watermark by requiring sampled sentences to pass fuzzy cluster-membership criteria, achieving TPR@1% above 97% with quality judged better than baselines by GPT-4o.

  12. SoK: On the Role and Future of AIGC Watermarking in the Era of Gen-AI

    cs.CR 2024-11 conditional novelty 5.0 of 10

    A systematization of AI-generated content watermarking that introduces a formal supply-chain definition and a property-based taxonomy.

  13. SEFD: Semantic-Enhanced Framework for Detecting LLM-Generated Text

    cs.CL 2024-11 conditional novelty 5.0 of 10

    SEFD combines retrieval-based semantic similarity with existing detectors and an adaptive pool to improve detection of paraphrased LLM-generated text in sequential streams.

  14. Towards Structurally Explainable Machine-Generated Text Detection: A Graph-Perspective Framework

    cs.CL 2025-05 reject novelty 4.0 of 10

    LM2OTIFS uses word co-occurrence graphs and GNNExplainer to detect and explain machine-generated text, with strong in-domain accuracy but unsupported faithfulness claims and a flawed theoretical proof.

  15. Survey on AI-Generated Media Detection: From Non-MLLM to MLLM

    cs.CV 2025-02 unverdicted novelty 3.0 of 10

    A survey organizing AI-generated media detection into Non-MLLM and MLLM based methods, with task and benchmark taxonomies.

Pith tools