mlsanityData to 7 Oct 2026

Watermark Smoothing Attacks against Language Models

arXiv 2407.14206 ↗First posted 19 Jul 2024cs.LGcited by 0

On the map: LLM safety and security → Watermarking generated content and models

Abstract (arXiv)

Watermarking is a key technique for detecting AI-generated text. In this work, we study its vulnerabilities and introduce the Smoothing Attack, a novel watermark removal method. By leveraging the relationship between the model's confidence and watermark detectability, our attack selectively smoothes the watermarked content, erasing watermark traces while preserving text quality. We validate our attack on open-source models ranging from 1.3B to 30B parameters on 10 different watermarks, demonstrating its effectiveness. Our findings expose critical weaknesses in existing watermarking schemes and highlight the need for stronger defenses.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Watermarking generated content and models, one of 12 directions in LLM safety and security. Builds on 37 of its 53 references that are in the map. First posted on arXiv 19 Jul 2024.

Builds on

37 of 53 references.

Closest ancestors · 23

Same direction, newest first

  1. 2024Bileve: Securing Text Provenance in Large Language Models Against Spoofing with Bi-level Signature4 Jun
  2. 2024Watermarking Language Models for Many Adaptive Users17 May
  3. 2024MarkLLM: An Open-Source Toolkit for LLM Watermarking16 May
  4. 2024A Statistical Framework of Watermarks for Large Language Models: Pivot, Detection Efficiency and Optimal Rules1 Apr
  5. 2024An Entropy-based Text Watermarking Detection Method20 Mar
  6. 2024Watermark Stealing in Large Language Models29 Feb
  7. 2024Can Watermarks Survive Translation? On the Cross-lingual Consistency of Text Watermark for Large Language Models21 Feb
  8. 2024GumbelSoft: Diversified Language Model Watermarking via the GumbelMax-trick20 Feb
  9. 2024Pseudorandom Error-Correcting Codes14 Feb
  10. 2023Towards Optimal Statistical Watermarking13 Dec
  11. 2023Mark My Words: Analyzing and Evaluating Language Model Watermarks1 Dec
  12. 2023Publicly-Detectable Watermarking for Language Models27 Oct
11 more
  1. 2023A Resilient and Accessible Distribution-Preserving Watermark for Large Language Models11 Oct
  2. 2023A Semantic Invariant Robust Watermark for Large Language Models10 Oct
  3. 2023Unbiased Watermark for Large Language Models22 Sep
  4. 2023Robust Distortion-free Watermarks for Language Models28 Jul
  5. 2023Provable Robust Watermarking for AI-Generated Text30 Jun
  6. 2023On the Reliability of Watermarks for Large Language Models7 Jun
  7. 2023Undetectable Watermarks for Language Models25 May
  8. 2023Who Wrote this Code? Watermarking for Code Generation24 May
  9. 2023Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense23 Mar
  10. 2023Can AI-Generated Text be Reliably Detected?17 Mar
  11. 2023A Watermark for Large Language Models24 Jan

Other references · 14

  1. 2020Language Models are Few-Shot Learnerscited by 32,199
  2. 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,491
  3. 2023Llama 2: Open Foundation and Fine-Tuned Chat Modelscited by 12,865
  4. 2022OPT: Open Pre-trained Transformer Language Modelscited by 3,464
  5. 2019The Curious Case of Neural Text Degenerationcited by 2,881
  6. 2018Hierarchical Neural Story Generationcited by 1,182
  7. 2022Contrastive Decoding: Open-ended Text Generation as Optimizationcited by 598
  8. 2024Qwen2-Audio Technical Reportcited by 543
  9. 2019Neural Text Generation with Unlikelihood Trainingcited by 489
  10. 2018Learning to Write with Cooperative Discriminatorscited by 119
  11. 2021Bad Characters: Imperceptible NLP Attackscited by 69
  12. 2021Paraphrastic Representations at Scalecited by 19
2 more
  1. 2021Fall of Giants: How popular text-based MLaaS fall against a simple evasion attackcited by 12
  2. 2024Edit Distance Robust Watermarks via Indexing Pseudorandom Codescited by 12

Built on by

Not cited yet.