Watermarking & Deepfake Detection
Cross-Family Generalization and Robustness of Text Watermarking Attacks and Defenses
Generated automatically from the limitations stated in 4 papers (ACL, ICML, ICLR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Existing evaluations of text watermark attacks, detection, and removal methods have almost exclusively tested narrow subsets of logit-based schemes (e.g., KGW and SIR) evaluated on fixed base models like OPT-1.3B. Because no single study has evaluated these attack and defense methods across fundamentally different watermark families—including sampling-based, dynamic-hash, content-adaptive, and error-correction-augmented schemes—practitioners cannot determine whether reported evasion and detection results reflect general properties or family-specific artifacts. Consequently, deployment decisions for provenance tracking in production language models rely on unverified theoretical compatibility claims rather than measured empirical robustness.
Why it matters
Provides the first definitive empirical map of which watermark attacks and defenses genuinely generalize across distinct algorithmic families versus which are restricted to logit-based implementations. This enables developers of provenance systems to choose watermark schemes with verified cross-family robustness profiles.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Benchmark existing watermark removal and evasion attacks (e.g., surrogate-based transfer attacks, perturbation oracles, and entropy-weighted detection) across four distinct text watermark families (logit-based KGW/SIR, sampling-based watermarks, dynamic-hash schemes, and adaptive watermarks) across multiple LLM backbones, measuring watermark bit recovery rate, detection $p$-value degradation, and output perplexity.
- 2
Evaluate cross-family attack transferability in settings where surrogate watermarks differ structurally from target watermarks (e.g., attacking a sampling-based or dynamic-hash target using logit-based surrogates), measuring transfer evasion success rates and detection AUC.
- 3
Quantify the empirical effectiveness of defensive error-correcting codes and adaptive hash keys against automated text perturbation attacks, measuring watermark retention thresholds under controlled semantic and edit-distance bounds.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
If transferability across disparate watermark families turns out to be uniformly high due to shared sensitivity to surface-level token edits, the evaluation will confirm baseline expectations without revealing distinct family-specific failure modes. The project could also be limited if emerging production watermarks remain entirely closed-source and cannot be accurately represented by open implementations.
Evidence
Each paper's own statement of the limitation, verbatim.
- Sandcastles in the Storm: Revisiting the (Im)possibility of Strong WatermarkingACL 2025
Only three watermarking schemes (KGW, SIR, Adaptive) and a fixed set of perturbation oracles were tested; other schemes and attack methods with error-correction may behave differently
- Revealing Weaknesses in Text Watermarking Through Self-Information Rewrite AttacksICML 2025
Evaluated only on watermarks for logit/sampling-based schemes generated by Opt-1.3B on C4; robustness against dynamic-hash or adaptive watermarks designed against this attack is untested
- A Transfer Attack to Image WatermarksICLR 2025
Effectiveness against a completely new target watermarking method substantially different from the surrogate families is unclear
- An Entropy-based Text Watermarking Detection MethodACL 2024
EWD is only implemented and evaluated on the KGW logit-based watermark; other watermark families (e.g., sampling-based) are only argued to be theoretically compatible
Nearest existing work
- An End-to-End Model for Logits-Based Large Language Models WatermarkingICML 2025
- Enhancing LLM Watermark Resilience Against Both Scrubbing and Spoofing AttacksNeurIPS 2025
- Bileve: Securing Text Provenance in Large Language Models Against Spoofing with Bi-level SignatureNeurIPS 2024
- An Ensemble Framework for Unbiased Language Model WatermarkingICLR 2026
- Discovering Spoofing Attempts on Language Model WatermarksICML 2025
- Waterfall: Scalable Framework for Robust Text Watermarking and Provenance for LLMsEMNLP 2024
- GumbelSoft: Diversified Language Model Watermarking via the GumbelMax-trickACL 2024
- Optimizing Adaptive Attacks against Watermarks for Language ModelsICML 2025
- Optimizing Watermarks for Large Language ModelsICML 2024
- Undetectable Watermarks for Language ModelsCOLT 2024
- XMark: Reliable Multi-Bit Watermarking for LLM-Generated TextsACL 2026
- On the Reliability of Watermarks for Large Language ModelsICLR 2024
- No Free Lunch in LLM Watermarking: Trade-offs in Watermarking Design ChoicesNeurIPS 2024
- Adaptive Text Watermark for Large Language ModelsICML 2024
- Watermarks in the Sand: Impossibility of Strong Watermarking for Language ModelsICML 2024
Related open problems
Watermarking & Deepfake Detection
Black-Box Verification and Trigger Generation for Model Watermarking and Detection
Current watermarking and deepfake detection methods require full white-box access to target model parameters and gradients during trigger construction or verification. When proprietary models are deployed solely behind inference APIs or distributed as encrypted binaries, these techniques cannot be applied at all. This leaves model owners and auditors unable to verify intellectual property theft, track provenance, or detect misuse across commercial black-box deployments.
Watermarking & Deepfake Detection
Black-Box and Model-Agnostic Deepfake and Watermark Verification for Closed-Source Generative Models
Current watermark verification and deepfake detection methods require white-box access to target model parameters, intermediate layer activations, or gradient information. Because leading generative models and vision-language systems are served exclusively behind closed commercial APIs, internal inspection methods cannot be executed by downstream verifiers or auditors. Consequently, defenders cannot detect deepfakes or verify watermarks when the generating model is proprietary, unknown, or inaccessible. Removing the precondition of internal weight and activation access is necessary for auditing real-world deployed models.
Watermarking & Deepfake Detection
Black-Box and Architecture-Agnostic Watermarking and Defense for Diffusion Models
Current proactive watermarking, attribution, and anti-customization defenses structurally assume white-box access to target latent diffusion models, requiring direct inspection of latent spaces, VAE fine-tuning, or gradient backpropagation through internal U-Net attention layers. In practical deployment scenarios, defenders and copyright holders must protect content against proprietary black-box APIs, pixel-space diffusion models, or non-VAE architectures where internal activations and weights are inaccessible. Because existing methods are tethered to specific LDM preconditions, there is currently no verified mechanism to provide provenance or anti-fine-tuning protection across arbitrary or black-box diffusion pipelines.
Watermarking & Deepfake Detection
Benchmarking Watermarking and Deepfake Detection Across Non-Latent and Flow-Matching Generative Architectures
Current watermarking and detection methods in generative media are almost exclusively developed and evaluated on standard latent diffusion models relying on Gaussian noise sampling. Consequently, it is unknown whether these techniques transfer, degrade, or fail entirely when applied to pixel-space diffusion, non-iterative architectures, or modern rectified-flow and flow-matching models such as SD3 and FLUX. As production generative pipelines shift away from standard latent diffusion, safety and provenance mechanisms risk operating under untested assumptions. A systematic evaluation across these architectural families is necessary to establish the empirical boundary of existing detection and watermarking schemes.