acceptodds
Withdrawn from ICLR 2027

Forensic Proxy Generation: Adapting Deepfake Detectors to Black-Box Generators

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

As commercial image generators rapidly evolve, deepfake detectors must continually adapt to the images they produce, typically through retraining or fine-tuning on newly generated images. However, collecting a large-scale dataset from each new generator can be impractical when only black-box access is available and query costs or usage restrictions limit extensive collection or downstream reuse. To this end, we introduce Forensic Proxy Generation (FPG), a two-stage framework that converts limited black-box access to a target generator into reusable proxy images for detector training. First, we treat the target generator as a teacher and distill teacher-relevant generative characteristics into a student generator. Second, we refine student-generated images through detector-response alignment. Under our problem setting, the resulting proxy images can be used for detector training while reducing the need for large-scale teacher querying and direct reuse of teacher-generated images. Across six target generators, FPG improves detection of teacher-generated images and balanced real–fake data without directly using teacher outputs for detector training. Under limited access, FPG expands a small number of teacher queries into proxy data that can outperform direct training on teacher outputs. These results show that FPG effectively adapts deepfake detectors to new generators under access and reuse constraints. The code is available at https://anonymous.4open.science/r/hci-code-3B8E.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
deepfake detection, generative models, generator distillation, media forensics

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9j84OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a method to improve the detection of blackbox generators with a limited number of synthetic images called Forensic Proxy Generation (FPG). It adapts a pretrained student generator using latent and frequency alignment with the teacher blackbox generator, as a way to generate proxy images using the student generator to increase data quantity. These proxy images are then used to train the detector. The method is evaluated using six target generators, with additional experiments using eleven detectors. It shows that the proxy datasets can improve detection over direct training with a few target images.

Critical strengths
  • Some of the experiments are interesting. Using only 10 or 25 target images to produce larger datasets that improve detection suggests that the method could be useful when collecting synthetic images is difficult.
  • The empirical evaluation covers multiple generators and detector architectures. The improvements across these settings show that the proxy images can be useful for adapting existing detectors.
Critical weaknesses
  • The paper does not show why this specific method is actually useful. Training a generator to mimic a target generator seems like a superfluous step instead of simply adapting detector training. The paper shows that expending the small dataset with the proxy generators can help, but it does not compare against a non-adapted student (does simply adding training data help the detector?) nor compare against a fine-tuned detector trained with the same number of target image (what is the degradation?). One hypothesis could be that the gains come from the student's pre-trained knowledge and the additional variation it provides. It is also unclear how the student can correctly mimic the target generator as we known from prior work that the model’s architecture plays a large role in the unseen artifacts a detector can pick up. These missing controls make it unclear whether training the student provides a meaningful benefit.
  • The evaluation is also restricted to face images. This is a useful setting, but it gives little evidence that the method would work on full images with different content, multiple faces, or partial edits, which I would argue are the much more common types of deepfakes. Since the method explicitly matches frequency characteristics, which is often much different on different image types, it is important to establish whether the improvements depend on the single-face portraits used in these experiments.
Questions for authors
  • Can you compare fine-tuning on 1,000 Nano Banana images with fine-tuning on 1,000 proxies, using the same detector initialization, augmentation, real-image data, and number of updates?
  • Can you report results using the pretrained student without teacher-guided adaptation, with the same prompts and image counts, both with and without Stage 2?
  • What augmentation and number of updates were used for direct training on the 10 or 25 target images in Figure 4? Were these matched to the proxy-training conditions?
  • Can you compare with simply reconstructing images through the student's VAE? This would help understand whether decoder artifacts explain some of the improvements.
Reviewer_UwTdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

FPG investigates how to improve the generalization of AI-generated image detectors when the target generator is accessible only as a black box and its outputs cannot be used directly for detector training. It proposes a two-stage framework: first, a student generator is trained via latent-space and frequency-domain alignment; then, proxy images are optimized by aligning a frozen detector’s responses to teacher and student images and used for downstream detector training. Experiments across six target generators and eleven detectors demonstrate improved detection performance and scalable proxy data generation under limited query budgets.

Critical strengths
  1. FPG addresses a practically relevant setting in which access to black-box generators is limited and teacher images cannot be directly used to train downstream detectors. It introduces a new approach to detector adaptation by converting limited teacher supervision into scalable proxy training data.

  2. The experiments cover six target generators and eleven detectors, evaluating cross-generator generalization, low query budgets, and parameter-efficient fine-tuning, providing broad empirical coverage.

  3. The method combines latent-space alignment, frequency-domain alignment, and detector-response alignment. Ablation studies investigate the roles of frequency constraints and second-stage optimization, providing preliminary empirical support for proxy image construction.

Critical weaknesses
  1. The student predicts a clean latent from a noisy source-image latent but is trained to align it with the latent of a different teacher-generated image. This may conflict with the original diffusion reconstruction objective. Please clarify the implementation and explain how these objectives are reconciled.

  2. The effectiveness of FPG is mainly demonstrated through final detection performance and average spectral similarity, without directly evaluating backbone representations. It remains unclear whether the gains stem from more generalizable representations or adjustments to the classification head and decision boundary. Please provide frozen-backbone linear probing, feature separability analysis, and ablations that separately update the classification head and backbone.

  3. In Stage 2, matching the predictions of a frozen detector on teacher-generated and proxy images may improve detector-specific adaptation rather than transfer forensic features from the target generator. Please compare Stage 1 only, random perturbations, and standard PGD targeting higher AI-generation scores under the same perturbation budget. Evaluate on independent detectors unseen during perturbation optimization to verify the contribution of Stage 2 and its cross-detector generalization.

  4. Although LoRA and Partial FT are evaluated, the advantage of FPG over simple few-shot adaptation remains unclear. Following the settings of FiSeR[1] and FSD[2], please freeze the backbone, fit the classification head using 10 labeled target-domain samples per class, and compare detection performance on the same test set.

[1] FiSeR: Fine-Grained Source Representations for Cross-Domain AI Image Detection. ICML 2026

[2] Few-Shot Learner Generalizes Across AI-Generated Image Detection. ICML 2025.

Questions for authors

Do improvements in detection accuracy at a fixed threshold reflect genuine gains in discriminative ability? The paper consistently uses 0.5 as the classification threshold, but the improvement in T-ACC is substantially larger than that in AUC in some experiments. For example, on NanoSet, T-ACC increases from 73.95% to 93.49%, while AUC increases only from 84.32% to 89.61%. The authors are encouraged to report the improvement in ACC at the optimal threshold.

The paper primarily compares FPG with Base and MDDA-inspired but lacks a comparison with PROBE [1], a recent method that uses detector-guided generation to construct challenging samples. The authors are encouraged to further discuss the differences between the two methods in sample construction and generalization, and to provide additional experiments or analysis under comparable settings.

[1] Where Detectors Fail: Probing Generative Space for Generalizable AI-Generated Image Detection. ICML 2026

Reviewer_Zv5eOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper considers a new problem of adapting deepfake detectors to a block-box generator, which can only be queried for a limited number of times, and the generated images cannot be directly used to train the detector. A two-stage framework termed Forensic Proxy Generation (FPG) is proposed to tackle this problem. Specifically, FPG distills the black-box generator by its limited output into a student proxy through frequency and latent alignment, and then adjusts the student outputs by adding bounded perturbations to improve detector-response alignment. Experiments suggest that training with the proxy data produced by FPG can effectively improve the detector’s performance.

Critical strengths
  1. The paper proposes a novel problem of adapting deepfake detectors to block-box generators under access and reuse constraints, which emphasizes the practical cost and feasibility of collecting a large number of outputs from the target generator for detector training.
  2. Experiments suggest that even with a small number of teacher queries (e.g., N=10/25), FPG can expands them into significantly larger proxy training sets, and the detector performance increases as the proxy data scale up, indicating the practical value of the proposed method.
Critical weaknesses
  1. The reuse constraint assumption may be impractical. Specifically, the problem setup in this paper assumes that the collected teacher outputs “may be used as limited supervision but not as downstream detector-training examples.” (Lines 119-121) The “limited supervision” here is not clarified. The proposed FPG utilizes the teacher outputs to train the student proxy via distillation, which can be regarded as reusing the data for “downstream model training”. Intuitively, it is unlikely that some providers prohibit detector training based on the generative content while permitting distillation.
Questions for authors
  1. Since the proposed method and the generators involved in the experiments are not limited to facial image generation, the proposed method can naturally be applied to general AI-generated image detection. What could be the potential challenges for this extension?

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

Related papers

Discussion (0)

Sign in to comment.