acceptodds
Withdrawn from ICLR 2027

Transferable Visual Identity Protection against Unauthorized Reuse in Closed-Source Image Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image-conditioned generation raises concerns about unauthorized use of visual identity. Adversarial image protection offers a lightweight defense against such misuse with small perturbations. However, existing defenses are typically developed on accessible generative pipelines, and their transfer across modern closed-source editors remains insufficiently characterized. To this end, we propose T-VIP, a transferable visual identity protection method for closed-source models that embeds a misleading identity signal into the protected image. Specifically, T-VIP constructs a transferable identity representation space from an ensemble of image encoders with lightweight projectors amplifying identity-discriminative cues. It then optimizes a push-and-pull misleading objective that drives the protected representation away from the source identity while keeping it within an identity-relevant manifold. Extensive experiments demonstrate that T-VIP substantially outperforms existing protection methods across diverse image generators without requiring white-box access. More importantly, it effectively transfers to closed-source systems, including GPT Image 2, Nano Banana 2, and Grok Imagine, highlighting its practicality for real-world visual identity protection.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Identity Protection, Image Generation

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_W65qOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper considers the problem of unauthorized use of visual identity in image-conditioned generation. The authors propose T-VIP, which combines frozen visual encoders with identity-supervised projection heads, then optimizes each image using a source-repelling objective and attraction toward an auxiliary identity bank. The method aims to achieve transferable identity protection against modern closed-source image editors without requiring access to target models. Experiments cover six open-weight and six closed-source generators, with substantial improvements over the evaluated transfer-based baselines.

Critical strengths
  1. Meaningful research topic. Protecting personal images against unauthorized identity reuse addresses an important privacy concern as reference-based generation becomes increasingly accessible.
  2. Strong empirical transfer. T-VIP substantially improves protection across diverse open-weight and closed-source generators without requiring target-model gradients.
  3. Simple and effective design. Identity-supervised projection heads strengthen the surrogate representation. The push–pull objective combines cosine-based separation from the source with soft nearest-neighbor attraction toward plausible alternative identity representations, with benefits supported by ablations.
  4. Complementarity across threats. Combining T-VIP with Anti-DreamBooth improves protection against training-based customization while retaining substantial protection against editing, supporting its relevance to scenarios involving both threats.
Critical weaknesses
  1. Insufficient evaluation of the source of transferability. The strong transferability of T-VIP may largely stem from generic ensemble-based adversarial transfer rather than its identity-specific design. Although the ablation study evaluates the contribution of encoder ensembling, it remains unclear whether comparable performance could be achieved by existing adversarial protection methods with similar ensemble strategies. The authors are encouraged to evaluate applicable baselines enhanced with multiple encoders, or compare against generic feature-space attacks using the same encoder ensemble, to better isolate the methodological contribution of T-VIP.
  2. Baseline configuration is unclear. Although the paper states a common (16/255) budget, it does not explain how IDProtector was adapted from its original (9/255) setting. It also omits IDProtector’s PGD variant without explanation, while T-VIP uses 300 per-image optimization steps, leaving optimization effort and method design partly confounded. More detailed baseline configurations and comparisons that account for optimization effort would help establish a fairer assessment.
  3. Protective perturbations are visibly noticeable. The protected images in the visualizations, particularly Figure 1, exhibit conspicuous perturbation patterns. Figure 6 acknowledges the protection–fidelity trade-off and reports better fidelity at smaller budgets, but this does not resolve the visual degradation associated with the default (16/255) setting used for the main results. This limits the strength of the high-fidelity and practical-deployment claims.
  4. Generalization across framing, pose, and distance is insufficiently evaluated. The examples predominantly show close-up portraits, and all ten evaluation prompts request head-and-shoulders portraits. Although the prompts include some gaze and expression changes, they do not establish effectiveness for substantial pose changes or full-body photographs. Random cropping during optimization does not substitute for evaluation under these conditions.
  5. The supplementary code lacks a README. The absence of a clear setup and execution guide makes reproduction unnecessarily difficult. Although the paper provides hyperparameters and pseudocode, running protection and evaluation are still needed.
Questions for authors
  1. How frequently does the pull objective steer representations toward particular real auxiliary identities? Do generated outputs resemble these identities, and does the mechanism reduce identity diversity across protected inputs?
  2. Could a practical feed-forward T-VIP variant reduce per-image runtime while retaining protection and transferability?
  3. How were the four ensemble encoders selected? Have alternative combinations or individual encoder contributions been evaluated?
Reviewer_8KdAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

T-VIP protects portraits against unauthorized identity reuse by adding a bounded perturbation before publication. It trains identity projection heads on frozen visual encoders and combines the projected and original features into an ensemble representation. For each image, it optimizes a perturbation that pushes its representation away from the source identity and pulls it toward embeddings in an auxiliary identity bank, without accessing the target generator. Evaluation covers two face datasets, six open-weight generators, and six closed-source services.

Critical strengths
  1. Substantial transfer to modern closed-source generators. The same protected image can be reused across target generators without re-optimization. On Nano Banana, T-VIP achieves a VIS of 0.082 versus 0.841 for IDGuardian, with a corresponding reduction in ISM at comparable per-image processing times. This is useful when users cannot inspect or control the downstream service.
  2. Ablations support the representation and objective design. Removing either the identity projection or the pull term degrades protection in Table 5. Reusing the original projection heads on CelebA-HQ also supports cross-dataset transfer.
Critical weaknesses
  1. Low VIS does not distinguish coherent identity changes from unrecognizable faces. Figure 8 assigns both a clearly different person and the absence of a recognizable face to the same low-score interval. The experiments do not report how often each occurs. Low VIS alone therefore does not establish that the outputs are coherent, recognizable portraits of a different person.
  2. Protection stability in practical use is not fully established. In Table 3, Gaussian blur increases Nano Banana’s VIS from 0.082 to 0.366, substantially weakening protection. Common operations such as resizing and resampling are not evaluated as post-processing steps. Only one output is generated per method–input–editor combination, leaving variation across repeated samples unmeasured. Claims about deployment reliability should be limited to the tested preprocessing and sampling conditions.
Questions for authors

Please address Critical Weaknesses 1–2.

Reviewer_zEXFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes T-VIP, a transferable adversarial protection method designed to prevent unauthorized facial identity reuse by closed-source image generation models. Unlike existing approaches that typically require access to specific generative architectures, T-VIP operates without knowledge of the target model or downstream editing instructions.

Critical strengths

This paper proposes T-VIP, a transferable adversarial protection method designed to prevent unauthorized facial identity reuse by closed-source image generation models. Unlike existing approaches that typically require access to specific generative architectures, T-VIP operates without knowledge of the target model or downstream editing instructions.

Critical weaknesses
  1. While the paper addresses a practically relevant problem and demonstrates strong empirical results, my primary concern is its limited technical novelty. The proposed T-VIP framework largely combines established techniques, including surrogate model ensembling, supervised contrastive learning, feature-space adversarial optimization, and input transformations. The push-and-pull objective also closely resembles conventional feature separation and target alignment strategies used in transferable adversarial attacks. Although these components are integrated effectively, the paper does not clearly establish what fundamentally new methodological insight is introduced beyond their combination. Furthermore, the ablation results suggest that the push objective accounts for most of the performance improvements, raising questions about the necessity and distinctiveness of the proposed misleading identity mechanism. Overall, the contribution appears primarily engineering-driven, with an emphasis on combining existing techniques and demonstrating their effectiveness across commercial image generators. While the extensive evaluation is valuable, the paper would benefit from a deeper investigation of the underlying transferability mechanism or a more substantial methodological contribution to better justify its significance for ICLR.

  2. The paper assumes that identity-discriminative features extracted by CLIP, OpenCLIP, and DINOv2 overlap with the identity-conditioning mechanisms used by modern image generators. Although the experimental results support the existence of transferability, the underlying mechanism remains insufficiently investigated. In particular, the proposed identity representation is constructed from discriminative image encoders, whereas the target generators may employ substantially different conditioning architectures. It is therefore unclear whether T-VIP exploits genuinely shared identity representations or more general vulnerabilities in visual conditioning.

  3. The push-and-pull objective is presented as a central methodological contribution. However, the ablation results in Table 5 indicate that the push term is responsible for most of the protection effectiveness. For example, removing the pull term increases VIS from 0.009 to only 0.055 on GPT Image 2 and from 0.082 to 0.123 on Nano Banana 2. By contrast, removing the push term leads to substantially larger performance degradation. These results suggest that the main mechanism may be source-identity feature disruption rather than misleading identity guidance toward alternative identities.

Questions for authors
  1. While the paper addresses a practically relevant problem and demonstrates strong empirical results, my primary concern is its limited technical novelty. The proposed T-VIP framework largely combines established techniques, including surrogate model ensembling, supervised contrastive learning, feature-space adversarial optimization, and input transformations. The push-and-pull objective also closely resembles conventional feature separation and target alignment strategies used in transferable adversarial attacks. Although these components are integrated effectively, the paper does not clearly establish what fundamentally new methodological insight is introduced beyond their combination. Furthermore, the ablation results suggest that the push objective accounts for most of the performance improvements, raising questions about the necessity and distinctiveness of the proposed misleading identity mechanism. Overall, the contribution appears primarily engineering-driven, with an emphasis on combining existing techniques and demonstrating their effectiveness across commercial image generators. While the extensive evaluation is valuable, the paper would benefit from a deeper investigation of the underlying transferability mechanism or a more substantial methodological contribution to better justify its significance for ICLR.

  2. The paper assumes that identity-discriminative features extracted by CLIP, OpenCLIP, and DINOv2 overlap with the identity-conditioning mechanisms used by modern image generators. Although the experimental results support the existence of transferability, the underlying mechanism remains insufficiently investigated. In particular, the proposed identity representation is constructed from discriminative image encoders, whereas the target generators may employ substantially different conditioning architectures. It is therefore unclear whether T-VIP exploits genuinely shared identity representations or more general vulnerabilities in visual conditioning.

  3. The push-and-pull objective is presented as a central methodological contribution. However, the ablation results in Table 5 indicate that the push term is responsible for most of the protection effectiveness. For example, removing the pull term increases VIS from 0.009 to only 0.055 on GPT Image 2 and from 0.082 to 0.123 on Nano Banana 2. By contrast, removing the push term leads to substantially larger performance degradation. These results suggest that the main mechanism may be source-identity feature disruption rather than misleading identity guidance toward alternative identities.

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

Related papers

Discussion (0)

Sign in to comment.