acceptodds
Under review as a conference paper at ICLR 2027

Prototype Margins and Classifier Decisions: A Geometric Analysis of Deepfake Attribution Robustness

Abstract

Deepfake attribution models are vulnerable to image degradations, but whether prototype geometry reliably reflects classifier decisions remains unclear. We introduce Attribution Margin Collapse (AMC), a diagnostic framework that examines class-center displacement and compares prototype margins with classifier logit margins. Two complementary statistics, AMC-S and AMC-F, summarize the mean prototype margin and the dataset-wide magnitude of negative prototype margins, respectively. Experiments on FaceForensics++ across four attribution models document robustness degradation, while JPEG analysis reveals similar aggregate trends in prototype and classifier margins. A complementary cross-domain study of binary deepfake detection on Celeb-DF reveals substantial disagreement: samples can retain positive prototype margins despite incorrect classifier predictions. These findings highlight both the utility and the limitations of prototype geometry for robustness analysis. AMC provides a framework for examining when geometric diagnostics reflect classifier behavior and when they offer an incomplete account of failure.

Primary area:
interpretability and explainable AI
Keywords:
Deepfake Attribution, Feature-Space Geometry, Prototype Margins, Classifier Decisions, Robustness Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_X9f8OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

Paper proposes a new metric, Attribution Margin Collapse (AMC), to examine the robustness of the prototype margin and its relation with the classifier prediction for deepfakes analysis

Critical strengths
  • The paper addresses an important topic to improve the robustness of deepfake attribution methods.
  • The current approach improves explainability and gives insights into deepfake-based analysis methods.
Critical weaknesses
  • The paper needs substantial work to improve its writing style. It is hard to follow, and the reader has to put substantial effort into understanding the underlying motivation and methods.

  • The novelty of the proposed method is limited. The approach of the prototype and the analysis proposed by the authors are well known in the field.

-The experiments conducted do not fully align with the motivation behind the proposed method. Specifically, the datasets considered primarily involve Gaussian blur and motion blur, which introduce relatively straightforward and easily distinguishable distortions. However, a more challenging and practically relevant setting involves deepfake images, where distinguishing manipulated content from authentic images is considerably more difficult. Evaluating the proposed method in such scenarios would provide stronger evidence of its effectiveness and practical applicability.

Questions for authors
  • How sensitive is the proposed method to the choice of the embedding?

-Can the authors provide results on more challenging datasets or deepfake techniques that generate images that are difficult to distinguish from real images?

  • Can the authors clearly state the assumption of Proposition 1
Reviewer_h64gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AMC, which compares prototype margins with classifier logit margins for deepfake attribution under image degradations. AMC-S and AMC-F summarize the mean margin and the negative part. On FF++ with four models, JPEG and blur hurt attribution, and prototype and logit margins drop together under JPEG. On Celeb-DF (binary), the two disagree a lot: 24% of samples have a positive prototype margin but a wrong classifier decision.

Critical strengths

I like that the paper doesn't just assume prototype geometry explains the classifier. Sections 3.1 and 5 keep the two decision rules apart, and Table 5 shows concretely where they disagree. That's a useful (partly negative) finding for anyone using prototype-based robustness analysis.

Critical weaknesses

1.The figures don't match the method or the tables. Section 3.2 says embeddings are L2-normalized and prototypes are class means, so distances should be at most 2 and margins within [-2, 2]. But Figure 4 shows distances around 20 to 25 and Figure 3 shows margins up to about 15. Figure 3 (clean) gives 78.1% positive margins while Table 3 gives 93.24%, and Figure 4 reports 68.0% accuracy under JPEG50 while Table 3 says 63.27%. Also, the Figure 3 histograms have nothing below 0 and a big spike at 0, which looks like clipping, so the margin "collapse" past the boundary isn't actually visible. I can't tell which computation the main claims are based on.

2.The cross-domain comparison mixes task and domain. Table 4 compares a 5-class FF++ model with a binary Celeb-DF model and reads the drop in correlation (0.90 to 0.52) as a domain shift effect. Section 5 itself says a fixed binary classifier is needed for this. As it stands, the drop could just as well come from the task change.

Questions for authors

1.Which features, model and samples were used for Figures 3 and 4? Are the features normalized there, and are margins clipped at 0? (clarification)

2.Can you run FF++ and Celeb-DF with the same binary model and the same source prototypes? (suggestion)

3.Section 4.4 calls Q2 an "early warning", but Q2 samples are already misclassified. Is there evidence that prototype margins fail later than logit margins as degradation gets stronger? (clarification)

4.Section 4.3 says the AMC trends hold for blur and across models, but I don't see any numbers. Could you add them? (clarification)

Reviewer_J3XzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Attribution Margin Collapse (AMC), a diagnostic framework examining the relationship between prototype geometry and classifier decisions in deepfake attribution under image degradation and domain shift. It uses fixed clean class prototypes to measure class-center displacement and prototype margins, summarized by AMC-S and AMC-F, and compares prototype margins with classifier logit margins. Experiments cover five-class attribution on FaceForensics++ and binary deepfake detection on Celeb-DF. Under JPEG compression, both margins exhibit similar aggregate degradation trends. On Celeb-DF, many samples retain positive prototype margins despite incorrect classifier predictions, highlighting the utility and limitations of prototype geometry for robustness analysis.

Critical strengths
  1. The paper explicitly distinguishes prototype geometry from actual classifier decisions by comparing correlation and sign agreement between the two margins, addressing a question with diagnostic value.
  2. The sample-level quadrant analysis provides information beyond accuracy. It distinguishes different failure modes of the nearest-prototype rule and the original classifier, offering insights into potential mismatches between the representations and the classification head.
Critical weaknesses
  1. Table 4 compares a five-class model on FaceForensics++ and a binarymodel on Celeb-DF.The decrease in correlation and increase in the Q2 proportion cannot be attributed solely to domain shift, as differences in the task, classification, and training settings also affect these statistics. The central coress-domain conclusion lacks a controlled comparison with the task and model settings held fixed.
  2. The proposition follows directly from the margin definition and does not establish a connection between class-center displacement and margin reduction.
Questions for authors
  1. Could you use the same binary classifier trained on the source dataset, with prototypes constructed solely from source-domain reference samples and held fixed throughout evaluation, and report the correlation and quadrant proportions on both datasets?
  2. How can the distance-increment condition in Proposition 1 be inferred from class-center displacement or verified experimentally? Does Corollary 1 require this condition to hold for every sample and every pair of adjacent perturbation levels? If not, what justifies the claimed monotonicity?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.