acceptodds
Under review as a conference paper at ICLR 2027

Latent Copyright Space for Model Provenance

Abstract

Model provenance and intellectual property protection are becoming increasingly important as trained models are treated as valuable digital assets and widely released, fine-tuned, and redistributed. A common approach is to embed ownership information into the model for subsequent verification. However, once a model is released and adapted to downstream tasks, the embedded ownership signals may be gradually weakened or even erased during fine-tuning. In this paper, we introduce a new concept, termed the latent copyright space, to capture such invariant characteristics of model derivation under downstream adaptation. Specifically, given a source model, we first construct a private low-dimensional space using a secret key. The source model is then minimally adjusted so that its projection into this space matches a prescribed copyright identity, with negligible effect on its original utility. After release, the model can be adapted using standard downstream training procedures. To verify a target model, we project it into the same copyright space and measure its consistency with the protected source. Models derived from the protected source are expected to preserve the encoded identity under downstream adaptation, whereas non-derived models exhibit lower identity consistency. Experimental results across vision and language models show that the proposed method can effectively distinguish models derived from a protected source from non-derived models under diverse downstream adaptation settings.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Model Provenance, Model Copyright Protection, Latent Copyright Space

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_kJaBOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces the Latent Copyright Space (LCS), a white-box method for verifying model provenance under downstream adaptation. Experiments on vision and language models demonstrate effectiveness across diverse downstream adaptation settings.

Critical strengths

The method is lightweight, requires no architectural modifications or triggers, and is independent of the downstream task. The small parameter perturbation introduced during watermark embedding has negligible impact on model utility.

Critical weaknesses
  • The paper is motivated by distillation, but LCS needs white-box access and the same parameterization, so it cannot detect distilled or re-architected models.
  • The newest vision baseline is from 2021. No SSL encoders are evaluated, and no SSL ownership methods are compared (SSLGuard, SSL-WM)
  • Unrealistic vision setup. Only ResNet-18 and ViT-Tiny on 10-class CIFAR subsets at 32×32.
  • Too few negatives to support the 0% FPR claim.
Questions for authors
  • Given the motivation of model distillation, can LCS detect distilled models or models with architectures different from the protected source?
  • How does LCS compare with more recent ownership verification methods, particularly those designed for self-supervised learning (e.g., SSLGuard and SSL-WM)?
  • Does LCS scale to larger pretrained models, such as ViT-B, fine-tuned on realistic datasets like ImageNet, with longer training schedules and different numbers of classes?
  • Can the authors report false positive rates on a larger, more diverse set of models, including ViT variants and independently fine-tuned copies of the same base checkpoint?
Reviewer_vLmuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Latent Copyright Space (LCS), a white-box framework for verifying whether a target model derives from a protected source model. LCS uses a secret key to construct a low-dimensional random projection of model parameters relative to a source checkpoint. A source identity is encoded through q-ary lattice symbols by optimizing a copyright loss alongside the original task loss. Verification decodes the projected target parameters and measures agreement with the registered identity. The paper provides a random-projection motivation and a sufficient condition for identity preservation. Experiments on ResNet-18, ViT-Tiny, and Qwen2.5-0.5B/1.5B report high verification rates under several downstream adaptation settings, with limited changes in task performance.

Critical strengths
  1. Simple and clearly specified protection mechanism.

LCS integrates identity encoding into a task-preserving optimization objective without changing the model architecture. Verification requires neither downstream training data nor adaptation history. This makes the approach understandable and potentially practical when white-box access is available.

  1. Encouraging results across several adaptation settings.

The experiments include full-parameter fine-tuning, LoRA, classification-head reinitialization, cross-dataset transfer, and successive adaptation. The language-model experiments also include negatives adapted from the corresponding unprotected checkpoints, which is a useful control for distinguishing the embedded identity from shared initialization.

Critical weaknesses
  1. The novelty and preservation mechanism need stronger justification.

LCS combines random projections, lattice quantization, and regularized watermark embedding, but its technical distinction from ST-DM remains unclear. Source centering can be absorbed into the lattice offsets. Moreover, the random-projection analysis assumes updates independent of the projection, which is not established for adaptation from the protected checkpoint. The theory therefore offers motivation and a conditional guarantee rather than demonstrating stability under realistic fine-tuning.

  1. The evaluation leaves comparative effectiveness and scalability uncertain.

Baseline thresholds and calibration procedures are insufficiently documented. Table 1(c), for example, reports baseline agreement scores above LCS’s threshold but zero TPR; different verification rules may explain this, but comparisons at matched false-positive rates are needed. Language-model experiments also cover only Qwen2.5-0.5B and 1.5B. These small checkpoints from one older family provide limited evidence for larger or newer models, especially without reporting the memory and runtime costs of constructing and applying the projection.

Questions for authors
  1. Method clarification: What is the precise technical distinction between LCS and ST-DM after accounting for source-dependent lattice offsets? Which component is responsible for the reported robustness improvement, and is there a controlled ablation supporting this attribution?

  2. Additional evidence that would affect my assessment: Can the authors provide results on a substantially larger checkpoint, such as a 7B/8B-scale model, and preferably a newer or different model family, under sustained fine-tuning? Such results would directly address the current uncertainty about scalability and generalization.

Reviewer_5JsKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • This paper introduces LCS, a framework for verifying model provenance under downstream adaptation. The motivation is that conventional ownership signals embedded in model parameters or representations may become unreliable after fine-tuning or other model modifications. To address this issue, LCS encodes source-specific copyright information in a private, low-dimensional subspace of the model parameter space.
  • Specifically, the method constructs a key-dependent orthonormal projection matrix and uses it to map parameter differences relative to the original source model into a low-dimensional copyright space. A source-specific identity is encoded using a qq-ary lattice-based quantization scheme. The source model is then optimized to align its projected representation with the target identity while preserving its original task performance. For provenance verification, a target model is projected into the same copyright space, and its decoded identity is compared with the registered identity using either exact or threshold-based soft verification.
  • The paper evaluates LCS on vision and language models under several downstream adaptation settings. The results suggest that the encoded identity can remain recoverable after the evaluated adaptations, with limited impact on downstream task performance. However, I have concerns about the method's novelty relative to existing parameter-based watermarking techniques, the calibration and statistical interpretation of false-positive rates, and the extent to which the experiments establish robustness to realistic model modifications and larger language models.
Critical strengths
  • The paper addresses an important model-provenance problem: verifying ownership or derivation when pretrained models are downloaded, fine-tuned, modified, and redistributed by downstream users. Developing verification mechanisms that remain effective after such modifications is practically meaningful.
  • The paper presents the latent copyright space, key-dependent projection, lattice-based identity encoding, and verification procedures in a systematic and generally understandable manner. The formulation provides a concrete framework for investigating identity preservation under parameter changes.
  • The reported downstream task performance is generally comparable to that of the corresponding unprotected models. The language-model perplexity results also suggest that copyright encoding does not substantially impair model utility in the evaluated settings. This is an important consideration for practical deployment.
Critical weaknesses
  • LCS constructs a low-dimensional random projection of model parameters and encodes ownership information through lattice-based quantization. However, projection- and quantization-based watermarking techniques have already been studied, including Spread-Transform Dither Modulation (ST-DM) [1]. Although the paper evaluates LCS against several existing baselines, the precise technical distinction between LCS and these approaches remains insufficiently clear. In particular, the authors should explain which aspects of the proposed formulation are fundamentally new and why these differences are expected to improve robustness under downstream adaptation. A more explicit comparison of the underlying objectives, encoding mechanisms, and robustness properties would help establish the contribution.
  • The paper discusses conventional trigger-based watermarking and several parameter- or representation-based approaches, but its experimental comparisons do not cover some recent methods relevant to model copyright protection, such as PLA [2], AGDI [3], and SIF [4]. These methods may differ from LCS in their verification mechanisms, threat models, or target model settings, so direct comparisons may not always be appropriate. Nevertheless, the paper would benefit from a clearer discussion of their differences and, where the settings are sufficiently comparable, additional empirical comparisons. This would help readers understand the scope of LCS and the scenarios in which it offers advantages over existing approaches.

[1] Spread-Transform Dither Modulation Watermarking of Deep Neural Network.

[2] Tracking the Copyright of Large Vision-Language Models through Parameter Learning Adversarial Images.

[3] Echoes of Ownership: Adversarial-Guided Dual Injection for Copyright Protection in MLLMs.

[4] SIF: Semantically In-Distribution Fingerprints for Large Vision-Language Models.

Questions for authors
  1. How does LCS scale to larger and more diverse models and tasks? The vision experiments focus on classification benchmarks, while the language-model experiments use only Qwen2.5-0.5B and Qwen2.5-1.5B on XSum and AG News. Although the paper evaluates several adaptation procedures, the current results provide limited evidence about scalability to larger language models and more diverse downstream applications. Could the authors provide additional experiments on larger models or more challenging settings? Even a focused evaluation with LoRA on a larger language model would help establish the practical applicability of LCS.

  2. How robust is LCS to non-training-based model modifications? The language-model experiments primarily consider fine-tuning and parameter-efficient adaptation. In practice, models may also undergo pruning, quantization, weight merging, or combinations of these operations. Have the authors evaluated whether the encoded identity remains verifiable under such modifications? Even a small controlled study would help clarify the robustness of LCS beyond the adaptation procedures currently considered.

  3. How is copyright identity verified when LoRA adapters are merged into the backbone model? The paper reports successful verification after LoRA adaptation, but it is important to clarify whether verification is performed on the frozen backbone together with separate adapters or on the full model after merging the adapter weights into the backbone. If the original backbone parameters remain unchanged during adaptation and verification, identity preservation may partly result from those unchanged parameters. Evaluating the merged model would provide stronger evidence about whether LCS remains effective when the model's actual parameter values are modified. Could the authors specify the verification protocol and analyze the effect of adapter merging?

  4. Could the authors report the computational cost of LCS? Section 6.1 identifies the GPU hardware used, but the runtime and resource overhead of copyright encoding and verification are not sufficiently characterized. Reporting encoding time, verification time, additional memory requirements, and the cost of downstream adaptation would help readers assess the practical overhead and scalability of LCS, particularly for larger language models.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.