acceptodds
Under review as a conference paper at ICLR 2027

DUNK: Dual-Space Non-Semantic Knowledge Consolidation for Continual Image Forgery Localization

Abstract

Image Forgery Localization (IFL) aims to locate manipulated regions at the pixel level. However, with the rapid evolution of generative artificial intelligence and image editing tools, newly emerging manipulation patterns continuously cause severe distribution shifts, making existing static IFL methods suffer from catastrophic forgetting when adapted to streaming forgery data. Therefore, in this paper, we focus on a challenging practical task called Continual Image Forgery Localization (CIFL), which requires continuously adapting to new manipulation patterns while preserving fine-grained localization ability for historical ones. To this end, we propose a Dual-Space non-semantic Knowledge Consolidation (DUNK) framework to maintain and consolidate historical authentic-forgery discrepancy knowledge in both spatial and frequency spaces. Specifically, a Dual-space authentic-forgery Prototype Memory (DPM) module is designed to compactly preserve structured authentic and forged prototypes without storing massive raw samples. Then, a Dual-space Adaptive Prototype Replay (DAPR) module adaptively coordinates spatial and frequency prototype replay according to their task-specific reliability, thereby preserving complementary historical forgery cues. Finally, a Multi-Perspective Knowledge Distillation (MPKD) module constrains the consistency between old and new models from mask, boundary, and feature perspectives to alleviate the degradation of fine-grained localization ability. Extensive experiments on multiple forgery localization benchmarks demonstrate the superiority of our method against existing IFL and continual learning methods in both localization accuracy and anti-forgetting ability. Our code will be released soon.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
computer vision, image forgery localization

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_otStOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies Continual Image Forgery Localization (CIFL), where a pixel-level forgery localizer learns sequentially from different manipulation datasets while retaining performance on previously encountered tasks without access to their training data. The authors argue that this setting requires preserving non-semantic discrepancies between authentic and forged regions, which existing continual semantic segmentation methods may not adequately capture. They propose DUNK, a PIM-based framework with three components. 1)DPM stores cluster means and diagonal variances of spatial features and frequency descriptors. 2)DAPR is designed to preserve historical knowledge by aligning the old and current models’ Mahalanobis-based prototype-relation distributions, with spatial and frequency losses weighted according to relation entropy and representation count. 3)MPKD further constrains changes in manipulation-mask predictions, boundary predictions, and features within informative regions.

Critical strengths
  1. A practically relevant continual-learning problem. This paper addresses the need to adapt image forensic models to emerging manipulation techniques while retaining localization performance on previously encountered ones. It focus on pixel-level localization without retaining historical raw images introduces meaningful constraints on both knowledge preservation and data storage, making CIFL a worthwhile setting for continual-learning research.

  2. An evaluation design that considers both localization performance and forgetting. The experiments cover five datasets and three task orders, with metrics assessing localization accuracy and historical-task retention. Adapting multiple continual-learning baselines to a common PIM backbone also helps isolate differences between knowledge-preservation strategies. Together with the component ablations, this design provides a useful basis for examining the method’s contributions and sensitivity to task order.

Critical weaknesses

Weakness 1: The central dual-space mechanism is not justified, and the frequency branch appears to provide no learning signal.

The main contribution of the paper is the claim that historical forgery knowledge should be consolidated jointly in spatial and frequency spaces. However, the frequency descriptor in Eq. 3 is a fixed, parameter-free function of the input image xix_i and the ground-truth mask yiy_i, and it never passes through the network. As a result, the old-model and new-model distances to the stored frequency prototypes are identical, the two relation distributions in Eq. 7 coincide, and the frequency replay loss LrepF=KL(rnF,t−1∥rnF,t)\mathcal{L}^F_{\mathrm{rep}} = \mathrm{KL}(r^{F,t-1}_n \Vert r^{F,t}_n) is identically zero with zero gradient. The reliability weight RFR_F is likewise constant with respect to the model parameters and only rescales the spatial replay term.

Despite this, Table 3 attributes a gain of more than 20 IoU points to the "FPP + DAPR" configuration over the base model. Unless a learnable frequency encoder has been left out of the description, this result cannot come from the method as stated. This undermines both the ablation study and the paper's core "dual-space" claim. In addition, the spectrum of a masked image F(g(x)⊙y)\mathcal{F}(g(x) \odot y) is largely determined by the shape and edges of the binary mask, so it is unclear why this descriptor should capture forensic traces rather than region geometry.

Weakness 2: Internally inconsistent results and insufficient experimental analysis to support the claimed advantages.

Under the paper's own definition of Average Forgetting (Eq. 53), the best historical score on a dataset equals its final score in Table 1 plus its forgetting in Table 2, and it cannot exceed 100. For the Adapter baseline, this gives implied peak scores of 109.5 F1 on AutoSplice, and 104.1 IoU / 110.5 F1 on MMTD-Set, which are impossible. Moreover, several entries report F1 values above the upper bound implied by the corresponding IoU — for example CoGaMiD on IMD2020 (21.9 IoU vs. 42.1 F1, bound ≈35.9\approx 35.9) and Joint-Train on MMTD-Set (79.8 IoU vs. 91.8 F1, bound ≈88.8\approx 88.8). This suggests either an undisclosed difference in how the two metrics are computed or errors in the tables. All results come from single runs without variance, while the reported margins over the strongest continual learning baseline are very large (around 19 IoU points), so these inconsistencies matter.

Beyond this, the experiments do not isolate the contributions of the key design choices. For example, the paper lacks:

  • an MPKD-only configuration, which is needed because MPKD is close to standard LwF-style output and feature distillation;
  • a comparison between the stored prototypes and a trivial or randomized memory, to show that the gains come from the prototypes rather than from old/new-model consistency alone.
Questions for authors
  1. The frequency prototype replay is presented as a core component of DUNK. Could the authors clarify whether there is any learnable component between the input image and the frequency descriptor in Eq. 3? If not, how can LrepF\mathcal{L}^F_{\mathrm{rep}} produce a non-zero gradient, and what explains the large gain of the "FPP + DAPR" row in Table 3? If there is one, please describe it fully: architecture, descriptor dimension, and how authentic-region descriptors are built.

  2. Tables 1 and 2 appear inconsistent under the paper's own definition of Average Forgetting. Could the authors explain exactly how IoU and F1 are computed (per-image or dataset-level, fixed or optimal threshold), account for the Adapter entries whose implied peak scores exceed 100, and provide the full accuracy matrices {ai,jm}\{a^m_{i,j}\} for all methods?

  3. Several datasets contain many authentic images with all-zero masks. Could the authors describe how these images are handled in training and in the IoU/F1 computation?

  4. The proposed DUNK framework combines prototype memory, adaptive prototype replay, and multi-perspective knowledge distillation, which appear to build on established continual learning techniques. Could the authors clarify what the core theoretical or methodological innovation of DUNK is beyond integrating these components?

I would be willing to raise my score if the authors can adequately address my concerns.

6 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/3)

[W1&Q1]: Frequency Replay Mechanism

Thank you for pointing this out. We further provide the details of the frequency replay mechanism in Eq. (6) of our latest manuscript(lines 270-278). Our frequency-based knowledge preservation branch provides learning signals through prediction-conditioned frequency prototype relation distillation. Specifically, Eq. (3) is used to extract frequency descriptors for constructing historical prototypes that preserve the distributional characteristics of authentic and forged regions. The resulting frequency descriptors are represented as 512-dimensional vectors. During frequency replay, we incorporate the current model's soft predictions into frequency extraction. Specifically, the descriptors are computed as:

vfgt=TF(g(xi)⊙yi⊙pt),v_{\mathrm{fg}}^t=\mathcal{T}_F(g(x_i)\odot y_i\odot p_t),

vautht=TF(g(xi)⊙(1−yi)⊙(1−pt)),v_{\mathrm{auth}}^t=\mathcal{T}_F(g(x_i)\odot (1-y_i)\odot(1-p_t)),

where TF\mathcal{T}_F denotes the fixed frequency descriptor extraction operator, yiy_i denotes the ground-truth manipulation mask, and ptp_t represents the current model's predicted forgery probability. As defined in Eqs. (7)–(9) in our latest manuscript, DAPR minimizes the KL divergence between the previous and current models' relation distributions over historical frequency prototypes. Consequently, the frequency replay loss can propagate gradients through the prediction-conditioned frequency descriptors to the prediction head φt(⋅)\varphi^t(\cdot), as follows:

∂LrepF∂θ=∂LrepF∂vt∂vt∂pt∂pt∂θ.\frac{\partial \mathcal{L}_{\mathrm{rep}}^{F}}{\partial \theta}=\frac{\partial \mathcal{L}_{\mathrm{rep}}^{F}}{\partial v^t}\frac{\partial v^t}{\partial p_t}\frac{\partial p_t}{\partial \theta}.

As shown in Tab. 3, Base+FPP+DAPR improves the average IoU/F1 from 32.5%/42.0% to 52.8%/64.6%, as preserving frequency-domain authentic–forgery discrepancy knowledge helps retain subtle manipulation traces, such as spectral perturbations and compression artifacts, that are difficult to capture in the spatial domain alone. Furthermore, by adaptively preserving authentic–forgery discrepancy knowledge across spatial and frequency domains, the combined SPP+FPP+DAPR configuration achieves 58.2%/70.6% in average IoU/F1, demonstrating the effectiveness of complementary dual-space knowledge in improving forgery localization performance.

We clarify the distinction between fixed frequency prototype construction and differentiable prediction-conditioned replay in the revised manuscript.

AuthorsOpenReview ↗

Official Comment by Authors (2/3)

[W2&Q2]: Evaluation Procedures And Additional Experiments

In Tab. 1, our reported image-level IoU and pixel-level F1 use different aggregation strategies. Specifically, IoU is calculated for each test image and then averaged across the dataset, ensuring that each image contributes equally despite variations in manipulated region sizes. In contrast, F1 is computed by aggregating pixel-level true positives (TP), false positives (FP), and false negatives (FN) across the entire dataset, following [1][2][3]. To ensure a consistent and transparent evaluation, we further report pixel-level IoU and F1. As below, DUNK achieves the best IoU/F1 of 59.7%/72.3%, leading the strongest baseline Adapter by 17.3%/15.2%, showing its effectiveness.

MethodCASIAv2IMD2020AutoSpliceMMTD-SetSID-SetAverage
Joint-Train59.6/74.762.5/76.986.2/92.684.8/91.882.8/90.675.2/85.3
PIM + LwF19.0/31.922.0/36.137.7/54.839.2/56.379.5/88.639.5/53.5
PIM + PLOP18.9/31.818.7/31.552.0/68.454.1/70.251.6/68.139.1/54.0
PIM + DKP19.5/32.721.1/34.846.7/63.755.8/71.660.5/75.440.7/55.6
PIM + Adapter20.6/34.220.0/33.456.4/72.160.4/75.354.4/70.542.4/57.1
PIM + CoGaMiD20.6/34.126.7/42.128.1/43.944.6/61.767.5/80.637.5/52.5
DUNK (Ours)36.1/53.130.6/46.978.3/87.875.7/86.277.9/87.659.7/72.3

[1] BusterNet: Detecting Copy-Move Image Forgery with Source/Target Localization [2] Multi-Supervised Encoder-Decoder for Image Forgery Localization [3] IMDL-BenCo: A Comprehensive Benchmark and Codebase for Image Manipulation Detection & Localization

Besides, we sincerely thank you for pointing out our oversight in Tab. 2 and we further provide the complete stage-wise accuracy matrices ai,jma_{i,j}^{m} for DUNK and the CL baselines below. We have carefully re-examined the corresponding results in the revised manuscript.

MethodDatasetCASIAv2IMD2020AutoSpliceMMTD-SetSID-Set
PIMCASIAv20.563/0.715————
IMD20200.163/0.2890.639/0.786———
AutoSplice0.116/0.2060.242/0.3270.877/0.919——
MMTD-Set0.068/0.1510.314/0.3420.288/0.4140.854/0.919—
SID-Set0.044/0.0910.157/0.2690.332/0.4280.426/0.5130.661/0.797
PIM+AdapterCASIAv20.756/0.850————
IMD20200.374/0.5530.564/0.715———
AutoSplice0.319/0.4090.214/0.3020.844/0.912——
MMTD-Set0.197/0.3490.207/0.3410.304/0.4520.894/0.915—
SID-Set0.184/0.3420.189/0.3340.564/0.7210.604/0.7530.545/0.705
PIM+CoGaMiDCASIAv20.571/0.627————
IMD20200.446/0.5660.468/0.694———
AutoSplice0.396/0.4670.351/0.5270.433/0.536——
MMTD-Set0.346/0.4020.279/0.4590.358/0.4770.743/0.811—
SID-Set0.217/0.3410.219/0.4210.308/0.4390.492/0.6170.687/0.806
DUNK (Ours)CASIAv20.734/0.812————
IMD20200.466/0.6790.541/0.718———
AutoSplice0.449/0.6470.429/0.5890.878/0.960——
MMTD-Set0.333/0.4660.383/0.5270.723/0.8360.887/0.940—
SID-Set0.387/0.5310.356/0.4690.792/0.8780.764/0.8620.751/0.876

We conduct an additional configuration, PIM+MPKD, which incorporates only the proposed multi-perspective knowledge distillation module without DPM and DAPR. We compare this configuration with PIM and PIM+LwF to distinguish the contribution of MPKD from conventional knowledge distillation. As below, PIM+MPKD outperforms PIM+LwF by 3.7%/4.9% in average IoU/F1, suggesting the benefits of jointly preserving mask, boundary, and feature representations.

MethodAvg. IoU (%)Avg. F1 (%)
PIM32.542.0
PIM+LwF40.953.5
PIM+MPKD44.658.4
DUNK (Ours)61.072.3

To further investigate the effectiveness of historical prototype construction, we consider two additional memory strategies: Random Feature Memory (RFM), which randomly samples historical features as memory representations, and Simple Mean Memory (SMM), which constructs memory representations through straightforward feature averaging. Both configurations preserve the same memory budget and knowledge distillation mechanism as DUNK. As below, our DUNK outperforms RFM+DAPR+MPKD and SMM+DAPR+MPKD by 8.2%/8.5% and 6.6%/6.1% in average IoU/F1, respectively, showing the effectiveness of our DPM in preserving authentic–forgery discrepancy knowledge in both spatial and frequency domains.

ConfigurationAvg. IoU (%)Avg. F1 (%)
RFM+DAPR+MPKD52.863.8
SMM+DAPR+MPKD54.466.2
DUNK (Ours)61.072.3
AuthorsOpenReview ↗

Official Comment by Authors (3/3)

[Q3]: Handling of Authentic Images

During training, authentic images are assigned all-zero ground-truth masks and are treated in the same manner as forged images, allowing the model to learn to distinguish authentic from manipulated regions.

[Q4]: Methodological Innovation

In this paper, we propose a Dual-Space Non-Semantic Knowledge Consolidation (DUNK) framework to address the Continual Image Forgery Localization (CIFL) task. Compared with existing continual learning (e.g., Continual Semantic Segmentation and Person Re-Identification) methods that primarily preserve semantic knowledge, the core methodological innovation of DUNK lies in preserving non-semantic authentic–forgery discrepancy knowledge to maintain historical forgery localization capabilities while adapting to new manipulation patterns.

Specifically, we introduce three complementary modules: (1) Dual-Space Authentic–Forgery Prototype Memory (DPM), which preserves historical authentic–forgery discrepancy knowledge through structured prototypes in both spatial and frequency domains, without storing historical images; (2) Dual-Space Adaptive Prototype Replay (DAPR), which adaptively replays historical prototype relations according to the reliability of spatial and frequency cues, thereby preserving complementary forgery knowledge across different tasks; and (3) Multi-Perspective Knowledge Distillation (MPKD), which maintains the consistency between old and new models from mask, boundary, and feature perspectives to preserve fine-grained forgery localization capabilities.

As shown in Tab. 1, our DUNK achieves 61.0%/72.3% in average IoU/F1, outperforming the strongest continual learning baseline, Adapter, by 19.3%/15.2%, respectively. As shown in Tab. 3, introducing SPP+FPP+DAPR improves the average IoU/F1 by 25.7%/28.6% over the baseline. Furthermore, incorporating MPKD brings an additional 2.8%/1.7% improvement in average IoU/F1, respectively. These results demonstrate the effectiveness of our proposed non-semantic knowledge consolidation framework for CIFL.

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

This paper studies continual image forgery localization, where a model sequentially learns from different forgery datasets without revisiting earlier images. DUNK combines spatial and frequency prototype memories, adaptive prototype-relation replay, and distillation of manipulation masks, boundaries, and features. Experiments on five sequential datasets report improvements in localization accuracy and forgetting over sequential fine-tuning and adapted continual learning baselines.

Critical strengths

Continual forgery localization is a practically relevant problem. The five-dataset evaluation and adaptations of several continual learning methods provide a useful empirical setting for studying it.

Critical weaknesses

1. Unclear frequency replay mechanism.

Eq. (3) defines a frequency descriptor independent of model parameters. Because the old and current models use the same descriptor and historical prototypes, Eqs. (6)–(8), as written, yield zero frequency replay loss and no gradient to the model. This cannot explain the average IoU/F1 increase from 32.5/42.0 to 52.8/64.6 for Base+FPP+DAPR (without SPP or MPKD) in Table 3. The trainable frequency pathway needs clarification.

2. Insufficient support for the non-semantic contribution.

Appendix C already adapts the baselines to forgery outputs, forged-region prototypes, or feature distillation with pseudo masks (PLOP). DUNK’s spatial memory stores Gaussian statistics of forged and authentic backbone features, without explicitly removing semantics. Fig. 1 illustrates PIM rather than CSS and does not establish semantic retention as the cause of forgetting. DUNK’s distinction from existing prototype replay and distillation remains unclear. Table 3 also lacks MPKD-only and fixed-weight replay controls.

Questions for authors
  1. Under Eqs. (3) and (6)–(8), how does frequency replay provide gradients to the model? Please specify the frequency representation actually used during training, and explain how it accounts for the gain of Base+FPP+DAPR in Table 3.

  2. What evidence shows that DUNK preserves less semantic content than the adapted baselines (e.g., object-category probing of the stored prototypes)? Separately, how does full DUNK compare with Base+MPKD and with dual-space replay using fixed weights?

2 comments
AuthorsOpenReview ↗

Official Comment by Authors

[W1&Q1]: Frequency Replay Mechanism

Thank you for pointing this out. We further provide the details of the frequency replay mechanism in Eq. (6) of our latest manuscript(lines 270-278). Our frequency-based knowledge preservation branch provides learning signals through prediction-conditioned frequency prototype relation distillation. Specifically, Eq. (3) is used to extract frequency descriptors for constructing historical prototypes that preserve the distributional characteristics of authentic and forged regions. During frequency replay, we use the current model's soft predictions to weight the authentic and forged image regions before frequency extraction, making the resulting descriptors differentiable with respect to the model's predictions. As defined in Eqs. (7)–(9) in our latest manuscript, DAPR minimizes the KL divergence between the previous and current models' relation distributions over historical frequency prototypes. Consequently, the frequency replay loss can propagate gradients through the prediction-conditioned frequency descriptors to the prediction head φt(⋅)\varphi^t(\cdot).

As shown in Tab. 3, Base+FPP+DAPR improves the average IoU/F1 from 32.5%/42.0% to 52.8%/64.6%, as preserving frequency-domain authentic–forgery discrepancy knowledge helps retain subtle manipulation traces, such as spectral perturbations and compression artifacts, that are difficult to capture in the spatial domain alone. Furthermore, by adaptively preserving authentic–forgery discrepancy knowledge across spatial and frequency domains, the combined SPP+FPP+DAPR configuration achieves 58.2%/70.6% in average IoU/F1, demonstrating the effectiveness of complementary dual-space knowledge in improving forgery localization performance.

[W2&Q2]: Non-Semantic Knowledge Preservation and Additional Experiments

Thank you for this insightful comment. Compared with existing continual learning methods such as Adapter, which may preserve semantic information entangled with historical forged-region features, our DUNK explicitly models authentic–forgery discrepancy knowledge in both spatial and frequency domains. As described in Eq. (65) of our latest manuscript, Adapter replays historical forged-region prototypes by enforcing their positive forgery responses, which contain object-level semantic information entangled with manipulation traces, potentially introducing semantic interference during continual adaptation. In contrast, our DUNK explicitly preserves both authentic and forged prototypes in spatial and frequency domains. As formulated in Eqs. (7)–(10), DUNK consolidates historical knowledge by maintaining prototype-relation distributions between the previous and current models. This discrepancy knowledge replay preserves historical relationships to authentic and forged prototypes. As shown in Tab. 1, DUNK outperforms Adapter by 19.3%/15.2% on average IoU/F1, demonstrating the effectiveness of preserving authentic–forgery discrepancy knowledge.

Moreover, we provide qualitative evidence of the distinction in the updated manuscript. As shown in Fig. 4, the manipulated region in Stage 3 corresponds to a dog, whereas in Stage 5, the dog is authentic and the manipulated region is a ball. Adapter incorrectly identifies substantial parts of the authentic dog as forged while failing to accurately localize the ball, suggesting potential interference from historical semantic information. In contrast, our DUNK accurately localizes the forged ball, illustrating its ability to better preserve authentic–forgery discrimination across different manipulation patterns.

We further compare DUNK with five fixed weighting configurations under identical experimental settings. As shown below, adaptive weighting outperforms equal weighting by 1.3%/0.6% in average IoU/F1 and achieves better performance than all fixed configurations, demonstrating the effectiveness of adaptively balancing spatial and frequency authentic-forgery discrepancy knowledge according to their reliability.

Weighting StrategySpatial / FrequencyAvg. IoU (%)Avg. F1 (%)
Fixed0.1 / 0.957.569.7
Fixed0.3 / 0.758.271.3
Fixed (Equal)0.5 / 0.559.771.7
Fixed0.7 / 0.359.170.9
Fixed0.9 / 0.158.971.0
OursDynamic61.072.3

We evaluate an additional configuration, PIM+MPKD, which incorporates only the proposed multi-perspective knowledge distillation module without DPM and DAPR. As shown below, PIM+MPKD outperforms the baseline by 12.1%/16.4% in average IoU/F1, suggesting the benefits of jointly preserving mask, boundary, and feature representations.

MethodAvg. IoU (%)Avg. F1 (%)
PIM32.542.0
PIM+MPKD44.658.4
DUNK (Ours)61.072.3
AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer ctLe,

We appreciate your time and insightful feedback on our work. As the rebuttal deadline is now only a few days away, we would like to respectfully ask whether our replies have resolved your questions. Please feel free to let us know if you have any further comments. We would be pleased to act on them promptly.

With thanks,

The Authors

Reviewer_15WdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies Continual Image Forgery Localization, where a model sequentially adapts to new manipulation patterns while retaining its ability to localize previously encountered forgeries. Built upon PIM, the proposed DUNK framework combines spatial and frequency authentic–forged prototype memories, adaptive prototype-relation replay, and distillation of masks, boundaries, and features. The authors evaluate the method on a sequence of five datasets and provide component ablations, distance-metric comparisons, and experiments with different task orders.

Critical strengths
  1. Forgery localization models must adapt to evolving image-editing techniques, while historical training data may not remain available. The continual learning setting therefore addresses a meaningful limitation of static forensic models.
  2. The paper compares several forgery localization methods and continual learning methods adapted to the same PIM architecture. Experiments with different task orders also provide a useful evaluation setting for studying continual forgery localization.
Critical weaknesses
  1. The frequency descriptor in Section 3.4 depends only on the input image and ground-truth mask, with no stated dependence on trainable parameters. It is therefore unclear how frequency replay supplies gradients or explains the substantial FPP+DAPR gains in Table 3.
  2. Several IoU/F1 results appear inconsistent with the stated metric definitions. For example, DUNK’s SID-Set scores in Table 1 exceed the F1 upper bound implied by its IoU when both metrics use the same predictions and averaging weights. The authors should clarify the evaluation procedures.
  3. Figure 5 shows a substantial first-stage performance gap between PIM and DUNK, although historical replay and distillation are not yet available. This unexplained difference raises concerns about whether the subsequent gains isolate the proposed continual learning mechanisms. Were initialization, data splits, and training settings identical?
  4. The ablations do not compare adaptive weighting with fixed equal weighting. Moreover, spatial and frequency representations are counted at different granularities, potentially making the weights reflect representation counts rather than reliability. Does adaptive weighting outperform equal weighting under matched settings, and how are the counts made comparable?
  5. The paper should discuss Learning Discriminative Noise Guidance for Image Forgery Detection and Localization and ForgeryGPT to clarify its relationship to noise-based and multimodal forgery localization. Threat in Frequency: Unveiling Hidden Attack Surfaces by Exploiting Stealthy Backdoors on Diffusion Models could additionally inform the discussion of frequency-cue robustness. How does DUNK differ from these approaches?
Questions for authors

Please refer to the Critical Weaknesses section above.

3 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

[W1]: Frequency Replay Mechanism

Thank you for pointing this out. We further provide the details of the frequency replay mechanism in Eq. (6) of our latest manuscript(lines 270-278). Our frequency-based knowledge preservation branch provides learning signals through prediction-conditioned frequency prototype relation distillation. Specifically, Eq. (3) is used to extract frequency descriptors for constructing historical prototypes that preserve the distributional characteristics of authentic and forged regions. During frequency replay, we use the current model's soft predictions to weight the authentic and forged image regions before frequency extraction, making the resulting descriptors differentiable with respect to the model's predictions. As defined in Eqs. (7)–(9) in our latest manuscript, DAPR minimizes the KL divergence between the previous and current models' relation distributions over historical frequency prototypes. Consequently, the frequency replay loss can propagate gradients through the prediction-conditioned frequency descriptors to the prediction head φt(⋅)\varphi^t(\cdot).

As shown in Tab. 3, Base+FPP+DAPR improves the average IoU/F1 from 32.5%/42.0% to 52.8%/64.6%, as preserving frequency-domain authentic–forgery discrepancy knowledge helps retain subtle manipulation traces, such as spectral perturbations and compression artifacts, that are difficult to capture in the spatial domain alone. Furthermore, by adaptively preserving authentic–forgery discrepancy knowledge across spatial and frequency domains, the combined SPP+FPP+DAPR configuration achieves 58.2%/70.6% in average IoU/F1, demonstrating the effectiveness of complementary dual-space knowledge in improving forgery localization performance.

[W2]: Evaluation Procedures

In Tab. 1, our reported image-level IoU and pixel-level F1 use different aggregation strategies. Specifically, IoU is calculated for each test image and then averaged across the dataset, ensuring that each image contributes equally despite variations in manipulated region sizes. In contrast, F1 is computed by aggregating pixel-level true positives (TP), false positives (FP), and false negatives (FN) across the entire dataset, following [1][2][3]. To ensure a consistent and transparent evaluation, we additionally report pixel-level IoU and F1. As shown below, DUNK achieves the best IoU/F1 of 59.7%/72.3%, outperforming the strongest baseline Adapter by 17.3%/15.2%, demonstrating the effectiveness of DUNK.

MethodCASIAv2IMD2020AutoSpliceMMTD-SetSID-SetAverage
Joint-Train59.6/74.762.5/76.986.2/92.684.8/91.882.8/90.675.2/85.3
PIM + LwF19.0/31.922.0/36.137.7/54.839.2/56.379.5/88.639.5/53.5
PIM + PLOP18.9/31.818.7/31.552.0/68.454.1/70.251.6/68.139.1/54.0
PIM + DKP19.5/32.721.1/34.846.7/63.755.8/71.660.5/75.440.7/55.6
PIM + Adapter20.6/34.220.0/33.456.4/72.160.4/75.354.4/70.542.4/57.1
PIM + CoGaMiD20.6/34.126.7/42.128.1/43.944.6/61.767.5/80.637.5/52.5
DUNK (Ours)36.1/53.130.6/46.978.3/87.875.7/86.277.9/87.659.7/72.3

[1] BusterNet: Detecting Copy-Move Image Forgery with Source/Target Localization [2] Multi-Supervised Encoder-Decoder for Image Forgery Localization [3] IMDL-BenCo: A Comprehensive Benchmark and Codebase for Image Manipulation Detection & Localization

[W3]: First-Stage Performance Improvement

Thank you for raising this important concern. In our method, we improve the boundary loss function of the baseline PIM. Specifically, the original PIM employs standard cross-entropy loss, where the large number of non-boundary pixels tends to dominate the optimization, leading to insufficient supervision of boundary regions. To address this issue, we introduce a dynamically class-weighted cross-entropy loss that adaptively adjusts the weights based on the proportion of boundary pixels in each batch, thereby alleviating class imbalance and improving boundary prediction. This modification contributes to the observed 17.1%/9.7% improvement in average IoU/F1 at the first stage. We will release our implementation code once accepted and explicitly clarify this implementation difference in the revised manuscript. Thank you again for your valuable comment.

AuthorsOpenReview ↗

Official Comment by Authors (2/2)

[W4]: Effectiveness of Adaptive Weighting

Thank you for this valuable suggestion. To evaluate the effectiveness of adaptive weighting, we further compare DUNK with five fixed weighting configurations under identical experimental settings. As shown below, adaptive weighting outperforms equal weighting by 1.3%/0.6% in average IoU/F1 and achieves better performance than all fixed configurations, demonstrating the effectiveness of adaptively balancing spatial and frequency authentic-forgery discrepancy knowledge according to their reliability.

In addition, we further evaluate an entropy-only adaptive weighting strategy to isolate the influence of representation counts. As shown below, entropy-only weighting outperforms equal weighting by 1.0%/0.4% and achieves performance closest to our full adaptive weighting strategy, suggesting that historical prototype-relation reliability provides effective guidance beyond representation counts.

Weighting StrategySpatial / FrequencyAvg. IoU (%)Avg. F1 (%)
Fixed0.1 / 0.957.569.7
Fixed0.3 / 0.758.271.3
Fixed (Equal)0.5 / 0.559.771.7
Fixed0.7 / 0.359.170.9
Fixed0.9 / 0.158.971.0
Entropy-only AdaptiveDynamic60.772.1
OursDynamic61.072.3

[W5]: Comparison with Related Works

Thank you for suggesting these relevant and insightful works. Noise Guidance exploits noise inconsistencies between authentic and forged regions to improve forgery localization, while ForgeryGPT integrates a Multimodal Large Language Model (MLLM) with a mask-aware forgery extractor to enable fine-grained localization and explainable forensic reasoning. Although these methods demonstrate strong performance in conventional forgery localization settings, they suffer from catastrophic forgetting when sequentially adapted to new data distributions. Different from these methods, our DUNK consolidates historical authentic–forgery discrepancy knowledge in both spatial and frequency domains, enabling continual adaptation to new manipulation patterns while preserving localization capabilities on previously learned manipulation patterns.

To further investigate this distinction, we further evaluate Noise Guidance and ForgeryGPT under the same five-task continual learning protocol as DUNK. As shown below, our DUNK outperforms Noise Guidance and ForgeryGPT by 25.9%/25.2% and 9.6%/9.0% respectively, verifying its effectiveness on the CIFL task.

MethodCASIAv2IMD2020AutoSpliceMMTD-SetSID-SetAverage
Noise Guidance8.9/16.113.5/23.631.7/47.547.1/63.674.3/84.635.1/47.1
ForgeryGPT40.7/49.229.6/40.357.3/72.262.4/76.267.0/78.851.4/63.3
DUNK (Ours)38.7/53.135.6/46.979.2/87.876.4/86.275.1/87.661.0/72.3

Thank you for highlighting Threat in Frequency, which investigates stealthy frequency-domain backdoor attacks against diffusion models. We explore incorporating insights from this work into DUNK to further enhance the robustness of frequency-domain knowledge preservation in future research.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer 15Wd,

We sincerely thank you for the time and careful consideration you have devoted to our manuscript. As the rebuttal phase is approaching its deadline, we would like to respectfully confirm whether our point-by-point responses have sufficiently addressed the questions you raised. If any residual concerns remain, we would be grateful for your further advice, and we remain ready to make additional revisions as needed.

Yours respectfully,

The Authors

Reviewer_q6iBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper tackles Continual Image Forgery Localization (CIFL), where models must adapt to sequentially emerging forgery methods without forgetting previously learned authentic–forgery discrimination. The authors propose DUNK, a Dual-Space Non-semantic Knowledge consolidation framework that preserves forgery cues in both spatial and frequency domains. DUNK combines a compact Dual-space authentic–forgery Prototype Memory (DPM) without raw-image replay, a Dual-space Adaptive Prototype Replay (DAPR) strategy that weights spatial and frequency replay by reliability, and a Multi-Perspective Knowledge Distillation (MPKD) module over mask, boundary, and feature levels. Experiments show state-of-the-art average IoU/F1 of 61.0%/72.3% and average forgetting of 18.5%/17.3%, outperforming IFL and continual-learning baselines.

Critical strengths

(1)Task-specific and storage-efficient knowledge preservation: DUNK explicitly targets non-semantic authentic–forgery discrepancy knowledge and stores compact dual-space prototypes instead of historical raw images, addressing privacy and storage concerns. (2)Complementary dual-space anti-forgetting design: DPM, DAPR, and MPKD jointly preserve spatial/frequency forensic cues and fine-grained localization behavior, adaptively balancing stability and plasticity during continual learning. (3)Strong and robust empirical validation: DUNK achieves clear SOTA gains in localization accuracy and anti-forgetting, remains effective under different task orders, and is supported by ablations, visualizations, and theoretical analysis.

Critical weaknesses

(1)The paper repeatedly emphasizes that the core of CIFL is to preserve “non-semantic knowledge,” i.e., category-independent discriminative knowledge between authentic and forged content. However, the method itself still performs K-Means clustering, prototype statistics, and relation distillation in the feature space, without demonstrating that these prototypes capture forgery traces rather than semantic content or image content. It lacks prototype visualization, semantic leakage analysis, and experiments examining whether frequency prototypes encode object shape or content. (2)DPM accumulates spatial/frequency prototypes on a per-task basis, but different forgery methods may produce similar features. Has the paper analyzed conflicts, redundancy, or negative transfer among cross-task prototypes? Has it considered prototype merging, pruning, task-similarity weighting, or forgetting mechanisms? (3)The paper does not report parameter count, training time, GPU memory usage, prototype storage, or inference overhead, nor does it provide a parameter/computation-matched comparison. It is therefore difficult to determine whether DUNK’s gains partly stem from larger capacity or additional training constraints. Although DPM avoids raw-sample replay, historical prototypes accumulate across tasks, and DAPR still computes distances between each token and historical prototypes; the actual storage and computation costs are not quantified. (4)The robustness of the model against attacks such as JPEG compression, Gaussian noise, and blurring has not been evaluated, which limits its practical application value. (5)Ωi\Omega_i includes high-confidence forged regions predicted by the old model. If the old model misses newly forged regions, those regions will not be constrained by distillation; if the old model falsely detects background, the errors may be passed on to the new model. Does this design introduce confirmation bias? Is there any handling of regions where the old model is uncertain? (6)DPM stores K_S + K_F prototype statistics per task, and DAPR computes the Mahalanobis distance between each token and historical prototypes. How do memory and computational complexity change as the number of tasks T increases? Have extension experiments with T > 5 been conducted?

Questions for authors

see weakness

4 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/3)

[W1]: Non-semantic Knowledge Preservation

Thank you for this insightful comment. Compared with existing continual learning methods such as Adapter, which typically preserve semantic information entangled with historical forged-region features, our DUNK explicitly models authentic–forgery discrepancy knowledge in both spatial and frequency domains according to the forged and real areas. Specifically, as described in Eq. (65) of our latest manuscript, existing continual learning methods merely replay historical forged-region prototypes by aggregating their positive forgery responses, which only contain object-level semantic information entangled with manipulation traces within forged regions, potentially introducing semantic interference during continual adaptation. In contrast, our DUNK explicitly preserves both authentic and forged prototypes in spatial and frequency domains. As formulated in Eqs. (7)–(10) of our latest manuscript, DUNK consolidates historical knowledge by accumulating historical discrepancy knowledge to authentic and forged prototypes represented by the relationships between these two kinds of prototypes. Therefore, as shown in Tab. 1, DUNK outperforms Adapter by 19.3%/15.2% on average IoU/F1, demonstrating the effectiveness of preserving authentic–forgery discrepancy knowledge.

Moreover, we provide qualitative evidence of the distinction in the updated manuscript. As shown in Fig. 4 in the latest manuscript, the manipulated region in Stage 3 corresponds to a dog, whereas in Stage 5, the dog is authentic and the manipulated region is a ball. Adapter incorrectly identifies substantial parts of the authentic dog as forged while failing to accurately localize the ball, suggesting potential interference from historical semantic information. In contrast, our DUNK accurately localizes the forged ball, illustrating its ability to better preserve authentic–forgery discrimination across different manipulation patterns.

In addition, we further compare three knowledge distillation strategies: naive Knowledge Distillation (KD) [1], Semantic Knowledge Distillation (SKD) [2], and our DUNK. As shown below, DUNK improves average IoU/F1 over KD by 15.3%/20.1% and SKD by 22.8%/23.2%, further demonstrating that preserving authentic–forgery discrepancy relations is more effective than conventional or semantic-oriented knowledge preservation in mitigating catastrophic forgetting.

MethodAvg. IoU (%)Avg. F1 (%)
PIM32.542.0
PIM + KD45.752.2
PIM + SKD38.249.1
DUNK (Ours)61.072.3

Here, non-semantic refers to the knowledge targeted by our consolidation mechanism, rather than a claim that the learned prototypes contain absolutely no semantic information. We have further clarified this distinction and discussed potential semantic leakage in the revised manuscript, particularly the influence of image content and region shape on frequency-domain prototypes.

[1] Distilling the Knowledge in a Neural Network [2] ACAM-KD: Adaptive and Cooperative Attention Masking for Knowledge Distillation

In our method, DUNK effectively mitigates cross-task prototype conflicts by preserving authentic–forgery discrepancy knowledge. In fact, cross-task prototype conflicts may arise when similar semantic content has different authenticity labels across tasks. For example, as shown in Fig. 1, a swan may appear as forged content in one task but as authentic content in another. Preserving semantic or identity-related knowledge (e.g., in the original designs of DKP, Adapter, and CoGaMiD) may intensify confusion between old and new forgery localization tasks and lead to catastrophic forgetting of historical forgery-trace knowledge. In contrast, DUNK constructs separate authentic–forgery prototypes in both spatial and frequency spaces and preserves their discriminative relations rather than relying on semantic identities. Moreover, DAPR adaptively balances spatial and frequency prototype replay according to their reliability, alleviating the influence of unreliable historical knowledge. These designs help mitigate semantic-related negative transfer during continual learning. As shown in Tab. 1, DUNK achieves the best average IoU/F1 scores of 61.0%/72.3%, demonstrating its effectiveness in preserving historical discriminative knowledge and mitigating catastrophic forgetting. -->

AuthorsOpenReview ↗

Official Comment by Authors (2/3)

[W2]: Cross-Task Prototype Conflicts

Thank you for your valuable comments. In our experiments, cross-task prototype conflicts may arise when similar semantic content has different authenticity labels across tasks. This is mainly because similar object features under different editing strategies may be aggregated in forged or real prototypes depending on whether they are forged or real, leading to conflicts during knowledge distillation.

To further illustrate this issue, we provide an additional qualitative comparison across different training stages in the updated manuscript. As shown in Fig. 4 in the latest manuscript, the manipulated region in Stage 3 is the river, whereas in Stage 5, the river is authentic and the manipulated region is the boat on the river. The baseline Adapter incorrectly identifies the authentic river as forged while failing to accurately localize the forged boat, suggesting that historical semantic information may interfere with new-task localization. Our method avoids the conflict and accumulation of erroneous knowledge by distilling the differentiated knowledge between different prototypes.

Furthermore, we compare the model performance under merging, pruning, task-similarity weighting, and forgetting 10% of historical prototypes, as shown in the table below. It can be seen that simply merging, pruning, or forgetting prototypes intensifies the confusion between old and new forgery localization tasks and leads to catastrophic forgetting of historical forgery-trace knowledge. While our method achieves optimal preservation of differentiated knowledge by adaptively balancing spatial and frequency prototype replay according to their reliability, alleviating the influence of unreliable historical knowledge.

Prototype Management StrategyAvg IoU (%)Avg F1 (%)
Prototype Merging (10%)58.769.8
Prototype Pruning (10%)57.968.9
Task-Similarity Weighting (10%)60.271.4
Random Forgetting (10%)56.868.1
DUNK (Ours)61.072.3

[W3]: Efficiency Comparison

We further provide a comprehensive efficiency comparison, including parameter count, training time, GPU memory usage, prototype storage, inference time, and FLOPs. All compared methods are implemented based on the same PIM backbone. As shown below, all methods have identical parameter counts (166.62M). Moreover, DUNK requires 6.34 h of total training time and 426.12 KB of prototype storage, compared with 6.68 h and 576.48 KB for CoGaMiD, while achieving substantial improvements of 22.5%/19.8% in average IoU/F1, as reported in Tab. 1 of the paper. These results demonstrate that DUNK achieves a favorable performance–efficiency trade-off while preserving historical authentic–forgery discrepancy knowledge.

MethodParams (M)Training Time (h)GPU Mem. (GB)Prototype Storage (KB)Inference (ms/img)FLOPs (T)
PIM166.625.6513.06054.554.23
+ LwF166.626.0413.67049.125.06
+ PLOP166.626.0613.88049.985.06
+ DKP166.626.2113.89771.8450.845.06
+ Adapter166.626.1914.12192.9651.125.06
+ CoGaMiD166.626.6813.86576.4850.745.06
DUNK (Ours)166.626.3414.40426.1252.535.06

[W4]: Robustness

Thank you for your suggestion. In our experiments, AutoSplice includes JPEG compression post-processing. We further introduce the image blurring subset of CoMoFoD (ELMAR 2013) and the Gaussian noise subset of FAU (TIFS 2012) to evaluate the robustness of DUNK under different image distortions. The training order is CASIAv2 → IMD2020 → AutoSplice → CoMoFoD → FAU.

As shown below, DUNK consistently outperforms existing methods, achieving the best average IoU/F1 of 68.5%/78.1%, surpassing the strongest baseline Adapter by 19.1%/16.6%. These results demonstrate the robustness of DUNK across diverse post-processing distortions.

MethodCASIAv2IMD2020AutoSpliceCoMoFoDFAUAverage
PIM8.2/15.216.5/27.037.5/49.648.9/61.069.0/80.236.0/46.6
PIM + DKP23.3/36.529.8/42.748.6/60.961.8/73.871.3/81.947.0/59.2
PIM + Adapter25.1/39.530.5/43.852.7/64.464.2/75.674.4/84.049.4/61.5
DUNK (Ours)44.8/58.746.2/59.877.3/86.085.2/92.089.1/94.168.5/78.1
AuthorsOpenReview ↗

Official Comment by Authors (3/3)

[W5]: Confirmation Bias

Thank you for pointing this out. Our DUNK mitigates potential confirmation bias by preserving historical authentic–forgery discrepancy knowledge rather than directly distilling the old model's forgery predictions. Specifically, DPM maintains both authentic and forged prototype distributions in spatial and frequency spaces. During continual learning, DAPR preserves the discriminative relations between these historical prototypes and current features, rather than relying on individual forged prototypes. Moreover, DAPR adaptively weights spatial and frequency prototype replay according to their reliability, reducing the influence of uncertain historical relations. These designs help mitigate potential error propagation while preserving historical forgery-discriminative knowledge.

[W6] Memory And Computational Overhead

Thank you for your valuable feedback. We further report the stage-wise training time, prototype storage memory and additional distance computation time of DAPR under the same RTX 4090 GPUs to evaluate the memory and computational overhead of DUNK as the number of tasks increases. As shown below, DUNK requires 6.34 h for all five training stages, compared with 6.68 h for CoGaMiD. Moreover, DUNK stores 426.12 KB of prototype statistics, compared with 576.48 KB for CoGaMiD, indicating a favorable balance between training efficiency and historical knowledge preservation.

MethodCASIAv2IMD2020AutoSpliceMMTD-SetSID-SetTotal Time (h)PSM (KB)
PIM1.191.570.391.261.245.650
+ LwF1.191.730.431.371.326.040
+ PLOP1.191.730.441.371.336.060
+ DKP1.211.760.511.391.346.21771.84
+ Adapter1.211.750.501.391.346.19192.96
+ CoGaMiD1.211.920.551.531.476.68576.48
DUNK (Ours)1.211.800.521.431.386.34426.12

In addition, we further measure the additional distance computation time of DAPR at each training stage, with the batch size set to 8. As shown below, the additional distance computation remains below 0.20 ms/batch throughout the evaluated five-task sequence, indicating that DUNK maintains manageable storage and computational overhead as historical knowledge accumulates.

Training StageCASIAv2IMD2020AutoSpliceMMTD-SetSID-Set
Distance Computation (ms/batch)00.100.100.150.19

To further evaluate the scalability of DUNK under a longer task sequence, we introduce two additional datasets, CoMoFoD (ELMAR 2013) and FAU (TIFS 2012), extending the benchmark from five to seven tasks. The new training order is CASIAv2 → IMD2020 → AutoSplice → CoMoFoD → MMTD-Set → FAU → SID-Set. As shown below, DUNK achieves the highest average IoU/F1 of 56.7%/69.8%, outperforming the strongest baseline Adapter by 10.6%/9.6%, demonstrating DUNK's scalability to longer task sequences.

MethodCASIAv2IMD2020AutoSpliceCoMoFoDMMTD-SetFAUSID-SetAverage
PIM4.3/7.910.9/19.331.8/47.738.2/54.638.1/54.756.8/71.863.2/77.034.8/47.6
PIM + DKP13.2/23.021.7/35.243.5/60.252.1/68.053.2/68.962.5/76.358.4/73.043.5/57.8
PIM + Adapter14.7/25.322.8/36.647.9/64.254.6/70.156.0/71.165.2/78.361.2/75.546.1/60.2
DUNK (Ours)26.8/41.832.6/48.665.9/78.965.8/78.763.4/75.372.0/83.270.2/81.956.7/69.8
AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer q6iB,

Thank you for your time and expertise in evaluating our work. With the rebuttal period concluding in several days, we wish to respectfully confirm whether our responses have adequately addressed your initial concerns. Any additional feedback would be greatly valued.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.