LineageMark: Multi-user White-box Watermarking for Contribution Tracing in Model Derivation Chains

summary

The gist

In open large language model (LLM) ecosystems, models are frequently adapted across multiple domains and applications, forming multi-stage derivation chains <ref:2606.17123#pg2>.

In short

LineageMark introduces a multi-user white-box watermarking framework to track contributions across complex, multi-stage model derivation chains in LLMs. It solves the limitation of existing methods by encoding watermarks using projection statistics over stable carriers, allowing multiple users to embed and verify their signals independently without needing access to the original model.

Key concepts

Stable Carriers
These are selected weight coordinates chosen because they are less sensitive to changes in the model parameters during derivation. They act as robust anchors for embedding watermarks, ensuring that subsequent updates or modifications only cause bounded statistical perturbations rather than complete signal loss.
Projection-Based Encoding
Instead of marking individual weights, LineageMark encodes each watermark bit as a sign-projection statistic calculated over a group of selected carriers. This method ensures that the embedding is resilient; even if some carriers change slightly, the overall statistical projection remains stable enough for accurate detection.
Key-Driven Verification
Each contributor uses their unique key to independently sample specific weight matrices and projection groups. This allows users to embed their watermarks and later verify them without needing access to the original model or other users' information, ensuring independent verification across the derivation chain.

Terminology used across episodes

This episode discusses

The paper

LineageMark: Multi-user White-box Watermarking for Contribution Tracing in Model Derivation Chains · Read on arXiv

Bingxue Zhang, Xiaofeng Xu, Feida Zhu

University of Shanghai for Science and Technology · Singapore Management University

In open large language model (LLM) ecosystems, models are frequently adapted across multiple domains and applications, forming multi-stage derivation chains. Consequently, tracking and verifying historical contributions is essential for model provenance and intellectual property protection. However, existing watermarking methods are mainly designed for single-user, one-time embeddings, often fail under repeated model derivation and incremental updates. To address this problem, we propose LineageMark, a multi-user white-box watermarking framework for model derivation chains. The framework encodes watermarks in model parameters using a projection-based approach. Stable carriers are first selected to reduce sensitivity to model changes, each watermark bit is then represented as a projection statistic over these carriers. Additional watermark insertions introduce only bounded perturbations in the projection space, and margin constraints are used to maintain signal integrity. We evaluate the effectiveness of LineageMark in multi-stage model derivation chains. Experimental results show that LineageMark preserves contributor watermarks across multi-stage derivation and supports incremental multi-user watermark insertion. Furthermore, it exhibits robustness against perturbations such as re-watermarking, fine-tuning, quantization, and pruning.

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "LineageMark: Multi-user White-box Watermarking for Contribution Tracing in Model Derivation Chains".

Elias: Here's the gist of the paper. In open large language model (LLM) ecosystems, models are frequently adapted across multiple domains and applications, forming multi-stage derivation chains <ref:2606.17123#pg2>.

Nadia: This paper proposes LineageMark to address the challenges of tracking and verifying historical contributions in these complex derivation chains by introducing a multi-user white-box watermarking framework <ref:2606.17123#pg5>.

Elias: Pruning redundant reasoning tokens The core problem addressed is that existing watermarking methods are mainly designed for single-user, one-time embeddings and often fail under repeated model derivation and incremental updates <ref:2606.17123#pg2>.

Nadia: LineageMark aims to verify which contributor watermarks are retained in a downstream derived model rather than recovering the full derivation topology <ref:2606.17123#pg5>.

Paper summary: Nadia: So, we're looking at this paper called "LineageMark: Multi-user White-box Watermarking for Contribution Tracing in Model Derivation Chains". Basically, the authors are tackling a big problem in the open large language model ecosystem where models get adapted across different domains and applications, creating these multi-stage derivation chains.

Elias: Right, and the core issue they're addressing is that most current watermarking methods only work for a single user embedding once. They fail when you have repeated model derivations or incremental updates to the model.

Nadia: So what LineageMark claims to do is propose this multi-user white-box watermarking framework specifically for those derivation chains, aiming to track and verify historical contributions so people can actually get some sort of model provenance and intellectual property protection.

Elias: The paper proposes encoding watermarks in the model parameters using a projection-based approach, selecting stable carriers first to make it less sensitive to changes, and then representing each watermark bit as a projection statistic over these carriers.

Priya: Before we get into the math, I want to know what this actually means for us on the ground. If we're looking at model derivation chains, what's the practical value of tracking who contributed what when?

Nadia: Well, it’s about verifying which specific contributor watermarks are still present in a downstream derived model instead of trying to recover the whole derivation topology. They claim this requires three things: lineage preservation, incremental embedding, and independent verification.

Elias: That independence is key because each contributor generates their coordinate groups and projection directions from private keys, meaning they can insert and verify their watermarks without needing the original model or any other users' information.

Priya: So if we trust this system to verify contributions independently, what's the actual risk that a derived model could accidentally lose or corrupt a watermark during those multiple derivation stages?

Nadia: The paper tackles that by using margin constraints to maintain signal integrity during these steps. They also use an empirical hidden-space Fisher matrix to select stable carriers, which is meant to reduce sensitivity to model changes and perturbation stability.

Elias: And they handle the incremental aspect by combining margin-aware embedding with redundant projection carriers, which is designed to preserve a decision margin for later stages and reduce interference from newly inserted watermarks.

Priya: That sounds like it’s trying to build a buffer against noise or accidental erasure during the process, but how robust is this against things like fine-tuning or quantization?

Nadia: The experiments show that LineageMark preserves contributor watermarks across multi-stage derivation and supports incremental multi-user watermark insertion. It also shows robustness against re-watermarking, fine-tuning, quantization, and pruning.

Elias: Specifically regarding those attacks, the paper mentions it maintains high extraction accuracy for both historical and newly inserted watermarks across derivation stages. It’s also robust against five parameter-efficient fine-tuning attacks: LoRA, Prefix-tuning, P-tuning, Prompt-tuning, and BitFit.

Priya: When you look at the results on robustness against those specific methods like LoRA or quantization to eight bits, how does that translate to real utility for someone who is just using a derived model?

Nadia: It means that even after those modifications—like quantization or pruning ratios up to zero point four—the framework still maintains a large gap between correct-key and invalid-key extraction scores, which keeps the false positive rate low under threshold testing.

Elias: The capacity of this system is governed by several things: the stable carrier pool size, key-driven sampling sparsity, vote redundancy, and false positive control under candidate-key verification. And interestingly, they found that stable carrier selection is actually the main computational cost.

Priya: So if we have to spend more on setting up these stable carriers than on the actual watermarking process itself, does that make it practical for widespread use?

Nadia: The authors point out that stable carrier selection dominates the end-to-end cost, suggesting a direction for future optimization where they might focus on making those carrier selections more efficient. But overall, it demonstrates the feasibility of moving from single-owner verification to verifiable multi-contributor provenance in open model ecosystems.

Elias: Ultimately, this work formalizes multi-user white-box watermarking in model derivation chains and designs a key-driven verification that lets each contributor embed and verify their watermarks independently without needing access to the original model or other users’ information.

Nadia: So, to sum up, LineageMark shows that you can track historical contributions reliably across repeated model changes while keeping the verification independent for each user. It maintains reliable watermark extraction under multi-stage derivation and incremental watermark insertion, even when facing common white-box modifications.

Priya: And from my side, it seems like the real win here is in establishing a verifiable chain of custody for model modifications within these complex ecosystems rather than just trying to find a single hidden mark.

Elias: Exactly. It moves the focus from just identifying a mark to proving where that mark came from in the process of making the model evolve. That’s what they are showing us with LineageMark: Multi-user White-box Watermarking for Contribution Tracing in Model Derivation Chains Bingxue Zhang, Xiaofeng Xu, Feida Zhu, on page one of that work <ref:2606.17123#pg1,LineageMark: Multi-user White-box Watermarking for Contribution Tracing in Model Derivation>.

Conclusion: Nadia: So, looking at the title and who wrote this paper, it’s "LineageMark: Multi-user White-box Watermarking for Contribution Tracing in Model Derivation Chains."

Elias: Right, it's by Bingxue Zhang and Xiaofeng Xu and Feida Zhu. It sounds like they’re trying to figure out how to track contributions without needing access to the original model or other users’ data.

Priya: So what does that mean in practice? If we can track provenance this way, what's the actual value for someone who just uses a derived model?

Nadia: It means you can actually verify which specific contributor watermarks are still there in the new model instead of trying to rebuild the whole history.

Elias: They’re using a projection-based approach with stable carriers and then margin constraints to keep the signal intact during those updates.

Priya: But they mentioned that carrier selection is actually their biggest computational cost, so if we want this running on real hardware, that’s a big hurdle for scalability.

Nadia: Exactly. And while it’s costly upfront, the paper shows it works reliably across different model sizes and even after things like quantization and pruning.

Elias: It proves you can embed watermarks independently for each user without needing the original model or any other user's information to do the verification.

Priya: So, if we look at this as a data privacy thing, it means we can potentially verify who contributed what while keeping that contributor information private.

Nadia: It’s about formalizing multi-user white-box watermarking in these derivation chains and showing that you can track historical contributors reliably.

Elias: It moves the focus from just finding one mark to proving where that mark came from in the process of making the model evolve.

More episodes

← Home