acceptodds
Under review as a conference paper at ICLR 2027

Protecting Images from Autoregressive Editing via Token Discrepancy Maximization

Abstract

Recent advances in autoregressive (AR) visual generation have broadened instruction-driven image editing capabilities, raising concerns about unauthorized image manipulation. Adversarial perturbations offer a proactive way to protect images from unauthorized editing before publication, but existing methods have largely been developed for diffusion-based editors. We study protection against autoregressive editors that represent source images as discrete visual tokens. In this setting, changing continuous image features may leave the selected tokens unchanged, limiting the effect of protection. We propose TokenShield, a quantization-aware method that encourages the protected image to select codewords farther from those selected for the original image. Its core objective, Ordered Codeword Competition, orders candidate codewords by their distance from the original selection and compares the nearest candidates on either side of each split. Minimizing the resulting distance gaps encourages farther codewords to become the nearest neighbors, while allowing the competing candidates to adapt during optimization. TokenShield generates perturbations through a frozen image tokenizer under a bounded pixel budget, without requiring editing instructions or autoregressive decoding. We evaluate TokenShield on eight image editors across two datasets, PIE-Bench and Emu Edit. The results show that TokenShield provides stronger protection than the compared methods by causing greater disruption to the edited images.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Protecting Images from Autoregressive Editing via Quantization-Aware Codeword Displacement Maximization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_EUwYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TokenShield, a quantization-aware adversarial protection method against unauthorized image editing by autoregressive (AR) generative models. The central observation is that perturbing continuous image representations does not necessarily change the discrete visual tokens consumed by AR editors, because hard quantization can map different continuous features to the same codeword.

Critical strengths
  1. The paper identifies an important limitation of conventional adversarial protection methods when applied to discrete-token image generators.
  2. The paper evaluates protection across eight editors and two datasets, including both AR and discrete diffusion architectures. This is a relatively comprehensive evaluation for a newly proposed protection setting.
Critical weaknesses
  1. Insufficient justification of the proposed quantization-aware protection objective.
  • Mismatch between codeword displacement and editing failure: The proposed method maximizes the Euclidean displacement between quantized codewords. However, codeword distance does not necessarily correspond to semantic changes or the failure of the requested editing operation. Consequently, it remains unclear whether maximizing the proposed objective directly improves the intended protection effectiveness.
  • Limited connection to target-editor representations: TokenShield optimizes perturbations using a single frozen LlamaGen tokenizer, while the evaluated editors employ heterogeneous visual representations. The paper does not establish whether the observed improvements result from disrupting the target editors' actual discrete token assignments or from more general adversarial feature transfer.
  • Limited guarantees under pixel-space constraints: Although OCC has desirable properties in the unconstrained feature space, these properties do not necessarily extend to bounded pixel-space optimization. The authors acknowledge that small OCC loss may not imply large hard-codeword displacement and that gradients can vanish or cancel during backpropagation. The practical implications of these limitations remain insufficiently investigated.
  1. Insufficient evidence of substantial practical improvements over strong baselines.
  • Marginal gains over direct PGD baselines: Although TokenShield substantially outperforms several external protection methods, its improvements over PGD-C and PGD-E are considerably smaller. For example, on Emu Edit, its average paired LPIPS improves only from 0.631 to 0.636 over PGD-E, and it is not consistently superior across individual editors. The absence of uncertainty estimates makes the reliability of these improvements difficult to assess.
  • Lack of direct editing-failure evaluation: The primary metrics measure differences between clean-input and protected-input editing outputs, rather than whether the requested edits are successfully prevented. The supplementary Target CLIP results show relatively small improvements over PGD-E (0.241 to 0.238 on PIE-Bench and 0.220 to 0.215 on Emu Edit). More direct instruction-following failure rates and source-content preservation evaluations are needed.
  • Limited robustness and computational analysis: Robustness is primarily evaluated under JPEG compression, without comprehensive comparisons against strong baselines under other transformations. Furthermore, TokenShield requires 500 PGD iterations, but the actual protection-generation cost and performance under matched computational budgets are not reported, making its practical advantage difficult to establish.
Questions for authors

See weakness and solve my concerns.

Reviewer_WwthOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies image protection for autoregressive image generation models with discrete tokenizers such as VQ. Existing continuous protection methods that mainly target VAEs do not work well in this setting, and the paper designs a protection method tailored to the quantized codebook. Through this, the proposed method achieves better performance than existing approaches.

Critical strengths
  • Most existing image protection methods are currently designed for diffusion models, and even then, many of them focus on somewhat classical and old models. This work targets AR-based editing methods with discrete tokens, so I think this is a very timely problem setting considering the recent trend of image generation models.

  • The paper investigates the performance and coverage of the proposed method across various settings and metrics.

Critical weaknesses
  • The robustness evaluation against purification is limited to JPEG, and there is no adaptive attack. Recently, several stronger purification methods have been proposed, and I think the paper should at least evaluate the proposed method against such approaches [1, 2].

  • Protection success is measured based on the distance from the clean edit. In most cases, if editing fails, it is reasonable to consider the protection successful, so I do not think this is necessarily a wrong evaluation. However, for example, simply shuffling the source tokens may also increase the distance regardless of whether the editing is actually disrupted in a meaningful way. Although [3] does not study AR-based models, I think its discussion on this issue may also be relevant here.

[1] Adversarial Perturbations Cannot Reliably Protect Artists From Generative AI, ICLR25

[2] Off-The-Shelf Image-to-Image Models Are All You Need To Defeat Image Protection Schemes, SaTML26

[3] Is Perturbation-Based Image Protection Disruptive to Image Editing, ICIP25

Questions for authors

Please see the weaknesses, I wonder how the model performs against the recent purification especially.

Reviewer_K79EOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates proactive image protection against unauthorized autoregressive image editing. The authors identify a mismatch between continuous perturbation objectives and hard-quantized visual tokens, and propose TokenShield, which uses Ordered Codeword Competition to maximize codeword displacement under a bounded pixel perturbation budget. The method requires neither editing instructions nor autoregressive decoding. Experiments on two datasets across eight image editors show that TokenShield achieves larger codeword displacement and greater disruption of editing outputs than existing methods.

Critical strengths

1.Image owners may want to share their images publicly while preventing others from manipulating them. This is a reasonable technical need, and exploring such protection mechanisms is valuable from a research perspective.

2.The mismatch between continuous objectives and hard quantization is clearly identified and theoretically supported.

3.The design of OCC has a clear mathematical connection to the original discrete objective, making the method relatively interpretable.

4.The experimental evaluation covers multiple image editors and two datasets. Comparisons with both existing protection methods and direct surrogate objectives (PGD-C and PGD-E) provide useful evidence for the effectiveness of the proposed optimization strategy. The additional analyses of image fidelity and JPEG robustness further strengthen the empirical evaluation.

Critical weaknesses

1.The main goal of this paper is to prevent unauthorized image editing, and the authors claim that TokenShield provides stronger protection than existing methods. Metrics such as Paired DINO, Paired LPIPS, and Hard-codeword Displacement can demonstrate that TokenShield changes discrete representations and produces larger differences in downstream editing outputs. However, I think that larger output differences do not necessarily indicate editing failure. For example, consider an image of a woman wearing red clothes, with an editing instruction to change the clothes from red to blue. The resulting image shows a man wearing blue clothes. In this case, the clothing color has been successfully changed, while the person's gender has also changed. Although the output difference is large, the requested edit is still successful.

2.In Figure 1, PGD-C, PGD-E, and TokenShield all achieve a 100% token flip rate. This indicates that, in the presented examples, conventional optimization objectives can already change all discrete tokens. Therefore, these results do not demonstrate that existing methods are actually limited by hard quantization or that this mismatch leads to insufficient protection effectiveness.

3.TokenShield shows clear advantages over traditional diffusion-based protection methods, but the most direct technical comparisons are actually PGD-C and PGD-E. The improvements of TokenShield over these baselines on downstream Paired LPIPS are relatively small.

Questions for authors

1.How do the authors ensure that larger Paired DINO and Paired LPIPS distances indicate actual editing failures rather than merely larger output differences? Could the authors provide a more direct evaluation of editing success rates?

2.In Figure 1, PGD-C, PGD-E, and TokenShield all achieve a 100% token flip rate. What evidence demonstrates that hard quantization mismatch is a bottleneck for existing methods and leads to insufficient protection effectiveness?

3.TokenShield shows relatively small improvements over PGD-C and PGD-E in downstream Paired LPIPS. What evidence demonstrates that OCC provides meaningful advantages over these simpler optimization objectives?

4.Could the authors provide a sensitivity analysis of the temperature parameter η in PGD-E to ensure a fair comparison?

5.Could the authors provide an ablation study on the selection and number of competition thresholds τ in OCC to demonstrate the necessity of aggregating multiple thresholds?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.