acceptodds
Under review as a conference paper at ICLR 2027

Evidence Is All You Need—But Not Forever: Open-World Continual Image Forensics

Abstract

Image forensics underpins multimedia security. Research is moving from detection under fixed conditions toward open-world forensics, where generators and image processing conditions continually change. Their joint effects can alter how visual features distinguish real from fake images, so past decision rules may become unreliable. Broader training coverage and continual updates improve detection. However, a theoretical gap remains: what should representations retain to reduce label requirements for future adaptation as decision rules change? To address this gap, we propose a new research framework and establish key theoretical guarantees. In a Gaussian model, we prove that coordinates learned from finite history reduce adaptation labels compared with retaining optimal past decision rules alone. These coordinates retain information that past decisions omit and future learning can use. Our bounds separately account for learning the coordinates from history and reaching a specified detection quality with current labels. We establish shared-space sufficiency, identifying when a shared space preserves full-space optimal AUROC as generation and processing change together. These results motivate Continual Updating of Evidence (CUE). With its visual encoder and initial detector fixed, CUE estimates within-class variation from historical and current statistics; current labels determine the class difference. One statistical objective yields the coordinates and classifier. We bound how statistical estimation error affects the classifier. Adding or subtracting window statistics supports continual learning and unlearning. Verified labels separately determine when fixed score contributions are suspended or restored. With 16 labels per class, CUE achieves state-of-the-art AUROC against 22 pretrained detectors across six standard benchmarks. CUE further turns accumulated experience into more data-efficient adaptation to new conditions. Together, these results establish a learning-theoretic foundation for open-world forensics.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
image forensics, AI-generated image detection, open-world learning, representation learning, continual learning, machine unlearning, few-shot adaptation, distribution shift

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 4.00 over 3 reviews, above 99% of the others.
Reviewer_aQUUOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies continual image forensics in the open world. Generators, compression methods, and propagation pipelines keep changing, so detection cues that worked in the past can stop working. A model then has to adapt to the current environment using only a few new labels. The central question is what information past experience should retain so that future labeling needs drop. In a Gaussian model, the authors prove that representations learned from history can retain information that the optimal past decisions left out, which reduces the labels required for the current task. They further analyze when a shared subspace keeps its discriminative power as generation and processing conditions change together. Building on this idea, the paper proposes CUE. The work offers a theoretical view of the long-term value of historical information, along with a workable method. The theoretical guarantees and the experimental results for the different implementations each hold within their own scope.

Critical strengths
  1. The problem matters. Detection cues that worked before can fail, and the paper directly asks what a model should keep and what it should update.
  2. The theoretical analysis is illuminating. Under a clear Gaussian setting, the authors prove that retaining historical information can help future adaptation more than retaining only the optimal past decisions.
  3. The statistical version of CUE is clearly designed. The coordinate transform and the classifier come from one objective, and updates leave the visual encoder untouched.
  4. The experiments give concrete support. The statistical version performs strongly on six benchmarks, and matched controls show that historical statistics improve few-shot adaptation.
Critical weaknesses
  1. Different experiments use different implementations, so a unified test of all continual learning abilities with one statistical version of CUE is still missing.
  2. In the six-benchmark comparison, CUE uses current labels while the other detectors do not, so the result is not a direct advantage at an equal supervision budget.
  3. The theoretical conclusions rest on specific assumptions, and the history reuse experiments cover few new generators, so wider applicability still needs verification.

I support acceptance. What draws me most is the analysis of which information in history deserves to be kept, and the attempt to turn that idea into a statistical update method.

Questions for authors
  1. Can the main continual learning experiments be run with one and the same statistical version of CUE?
  2. Does the CUE advantage survive once the strongest baseline also receives 16 labels per class?
  3. When the historical and current distributions differ a lot, can the method lower the weight of history automatically?
  4. Are the gains from history reuse stable when the set of historical generators changes?
  5. With only one or two labels per class, is class mean estimation the main bottleneck?
  6. How much do a few incorrect audit labels affect the update?
  7. After removing a window, can you report an error comparison against rebuilding the model from the remaining data alone?
  8. Is the classifier selection stable for single-image or small-batch prediction?
Reviewer_A9CAOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes CUE for image forensics under changing generators and processing conditions. With the visual encoder and initial detector fixed, CUE uses historical and current statistics to estimate within-class variation, while current labels determine the class difference. The coordinates and classifier are obtained from one statistical objective. Adding or subtracting window statistics supports continual learning and unlearning, and a separate controller suspends or restores fixed score contributions.

Critical strengths
  1. The problem is important. In practical image forensics, a feature that is useful under one generation or processing condition may become unreliable under another. Learning what information should be retained for future adaptation is a meaningful direction beyond improving detection on fixed benchmarks.
  2. The paper provides theoretical justification for retaining information beyond past decision rules. In the Gaussian construction, finite historical observations preserve a direction omitted by optimal past decisions and reduce the required current labels. The shared-space analysis also explains when a low-dimensional representation can preserve optimal AUROC under joint generation and processing changes.
  3. The proposed statistical update is intuitive and efficient. The same objective produces the coordinates and classifier, and retained windows can be removed by subtracting their statistics. The classifier stability bound and exact-arithmetic reconstruction guarantee provide useful support for the design.
  4. Extensive experiments cover detection, few-shot adaptation, continual retention, and reversible contribution use. The joint-moment CUE comparison with 16 labels per class outperforms the 22 pretrained detectors on shared query subsets of six benchmarks. The contribution-control experiments also show that suspending harmful contributions and restoring them when conditions return can be better than keeping them active or deleting them permanently.
Critical weaknesses
  1. The experiments use several different implementations, including joint-moment CUE, neural adaptation, a generalized-eigenvector method, and a stored-classifier reference. Appendix A clarifies these differences, but the main text is still difficult to follow. The paper should make clearer which results support the proposed statistical objective and which are reference results.
  2. The theoretical label advantage is established in a constructed Gaussian setting with substantial historical data. It does not directly imply the same label reduction for image features. The matched experiment is useful, but further evaluation across more conditions and feature encoders would strengthen this conclusion.
  3. The audit window needs to represent the query distribution. Appendix I shows a substantial performance drop when support from FLUX is used for SD3.5 queries. Handling mixed or mismatched windows remains important for practical deployment.
Questions for authors
  1. Could the authors provide a compact table mapping each experiment to its implementation, feature encoder, and update rule?
  2. How should the historical windows be selected or removed when their statistics become unreliable for the current condition?
  3. How would CUE detect a mismatch between the audit window and query distribution without generator or processing identities?
Reviewer_TuvoOpenReview ↗
Rating
4(+2)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates open-world continual image forensics under evolving generation and processing conditions. It proposes Continual Updating of Evidence (CUE), which leverages historical within-class statistics and current labeled samples to adapt forensic decision rules while keeping the visual encoder frozen. The method employs covariance-based coordinate calibration, closed-form classifier updates, and reversible activation of fixed forensic contributions.

Critical strengths
  1. The paper addresses an important practical challenge: forensic evidence may lose discriminative value as generators and image-processing conditions change jointly. The distinction between observable forensic traces and their changing decision utility is particularly interesting.

  2. The Gaussian construction demonstrates that historical observations can preserve information omitted by optimal historical decision rules. The analysis also provides conditions under which a shared feature subspace maintains optimal discrimination across changing conditions.

  3. CUE derives a closed-form classifier from historical covariance statistics and current class differences. Maintaining sufficient statistics enables incremental updates and window-level removal without storing historical images.

Critical weaknesses
  1. Inconsistency between the proposed method and experimental implementations. The primary CUE formulation is the joint-moment objective in Eq. (4). However, several major experiments employ different learning procedures. Specifically, Table 1 uses Community Forensics prototype/kernel adaptation, Figure 5 uses IAPL/CLIP neural adaptation, and Table 2 evaluates a stored-classifier reference. Table 3 further relies on a generalized-eigenvector reference. Although these differences are disclosed in Appendix A, the main results do not consistently validate the proposed Eq. (4) algorithm, weakening the overall experimental evidence.

  2. The joint-moment objective in Eq. (4) closely resembles covariance-regularized residual linear regression on frozen features. The coordinate transformation in Eq. (5) essentially whitens the regularized within-class covariance. When the correction subspace is unrestricted ((R=I)), this transformation is full-rank and does not explicitly discover or preserve a new low-dimensional forensic representation. The methodological distinction from conventional covariance-aware regression or discriminant analysis remains insufficiently justified.

  3. Proposition 1 establishes a substantial adaptation-label advantage in a constructed Gaussian setting, but its augmented-coordinate representation differs from the joint-moment learner in Section 4. Similarly, the shared-space identification results concern another generalized-eigenvector construction. Proposition 2 directly analyzes Eq. (4), but only bounds classifier sensitivity to covariance estimation errors while fixing current class means. Therefore, the theoretical results do not directly establish the claimed label-efficiency advantages of the implemented CUE algorithm.

  4. The six-benchmark comparisons provide CUE with 16 labeled samples per class, whereas the 22 pretrained detectors receive no target-specific supervision. Moreover, the method benefits from a strong pretrained Community Forensics backbone and additional initialization and calibration data. Comparisons against conventional few-shot linear adaptation methods using identical frozen features, support samples, and tuning budgets are necessary to isolate the actual contribution of CUE.

  5. The matched historical-reuse experiment reports an average improvement of 1.66 percentage points in normalized AUROC–label-budget area using 256 additional historical labels, but covers only four conditions. Furthermore, continual performance partly relies on stored-classifier selection using unlabeled batch descriptors, which may be sensitive to mixed distributions and batch composition. The reversible evidence controller also supports only activation or suspension of fixed contributions, limiting its ability to handle changes in evidence polarity.

Questions for authors
  1. Can the authors evaluate the exact joint-moment CUE formulation in Eq. (4) on the main few-shot and continual-learning benchmarks? How do its results compare with the different reference implementations reported in Figures 5 and Tables 1–3?

  2. Since the coordinate transformation is an invertible reparameterization when (R=I), what is the fundamental advantage over regularized linear regression, shrinkage LDA, or covariance-preconditioned linear adaptation? Could the authors provide controlled comparisons using identical frozen representations?

  3. Can the authors clarify which theoretical guarantees directly apply to the practical joint-moment learner? In particular, is it possible to establish an adaptation-label or AUROC guarantee for Eq. (4) that accounts for both historical covariance estimation and current class-mean estimation?

  4. How does performance change with increasing historical data, irrelevant historical distributions, or contradictory forensic evidence? Can the authors provide a more systematic analysis of historical label costs and the conditions under which accumulated statistics improve or degrade adaptation?

  5. How sensitive is the stored-classifier routing mechanism to query-batch size, class imbalance, and mixed-world inputs? Additionally, can the reversible evidence controller handle cases where a previously useful forensic contribution becomes useful again with the opposite scoring orientation?

Comments on the paper

2 comments

Dear reviewers,

As the discussion period approached the end, please take a moment to read the authors’ responses and share your thoughts on any remaining points of discussion. If the discussion has changed your assessment, please update your review accordingly.

Thank you for your time and contributions!

AuthorsOpenReview ↗

Discussion Summary for the Area Chair and Reviewers

Dear Area Chair and Reviewers,

We thank the Area Chair and reviewers for their careful reading and constructive discussion. We would like to summarize the central result of the paper and the additional evidence provided during the discussion.

Our goal is to advance a fundamental question in open-world forensics: when generators and image-processing conditions continue to change, what should a detector retain from its history so that future adaptation becomes less label-intensive? Our theory shows that optimal past decisions need not retain all information useful for future learning. In an explicit Gaussian model, finite history can preserve directions discarded by past decisions but needed for future class separation, reducing the number of current labels required for adaptation. We also characterize when a shared space preserves optimal discrimination under joint changes in generation and processing. This gives a theoretical basis for maintaining reusable forensic evidence across changing conditions.

CUE is the statistical realization of this idea. It uses historical within-class statistics and current labels to update forensic coordinates and the classifier, while supporting continual addition and exact window removal. During the discussion, we evaluated the same Eq. (4) learner with the same frozen features, current supports, and shared source-development budget on six detection benchmarks, a six-scene few-shot benchmark, and two 12-domain continual streams. Under equal-label frozen-feature controls, Eq. (4) reaches 99.21% macro AUROC on the six detection benchmarks, compared with 98.92% for OAS-LDA and 98.88% for ridge. On the few-shot benchmark, it reaches 96.27%, compared with 96.03% and 95.59%.

The revision also separates direct Eq. (4) evidence from neural adaptation, stored-classifier, and generalized-eigenvector reference results. The additional analyses clarify the conditions for historical gains, theoretical guarantees, window removal, and batch-dependent routing. Anonymous code, checkpoints, reproduction materials, and the revised manuscript are available with the submission for checking the reported results and their scope.

We are sincerely grateful to the reviewers for their thoughtful questions and constructive discussion. Their feedback helped us make the connection between the theory, the statistical learner, and the empirical evidence more precise.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.