acceptodds
Withdrawn from ICLR 2027

SyRC: Synthesis–Resampling Consistency for Deepfake Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Online forgery synthesis diversifies deepfake training data, yet detectors trained on these examples can remain sensitive to downsampling and restoration. To complement forgery diversity with stability under resampling, we introduce Synthesis–Resampling Consistency (SyRC). During training, SyRC refreshes self-blended forgeries online and constrains the feature discrepancy between each augmented instance and its resampled view. By sharing the underlying image and base augmentation draw, the two views isolate resampling-induced changes for the consistency constraint. The additional operations are confined to training, preserving single-view inference. With GenD trained on FaceForensics++, SyRC improves detection on four target datasets. The mean video-level area under the receiver operating characteristic curve (AUC) increases from to . Matched comparisons also favor shared-base pairing over independent augmentation, supporting controlled within-instance consistency as a useful complement to synthetic forgery diversity.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Deepfake Detection, Cross-Dataset Generalization, Online Forgery Synthesis, Resampling Consistency, Feature Consistency Learning

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

SyRC combines online self-blended forgery synthesis with consistency under downsampling and restoration. During training, selected fake entries are replaced with newly generated SBI samples. Each real or fake instance is augmented once, and a paired view is obtained by downsampling to 128 pixels and restoring the original size. A squared-cosine feature loss couples the views, while one randomly selected view receives the detector's original task supervision. Experiments with GenD and Effort report improved cross-dataset detection, supported by component ablations and source-validation stability diagnostics. Inference remains unchanged.

Critical strengths
  1. The training procedure is simple and preserves single-view inference without introducing additional inference modules. Sharing the base augmentation provides an interpretable way to constrain changes associated with a specific transformation.

  2. The internal controls are useful: the paper separates online synthesis, resampling exposure, and paired consistency, and compares shared-base with independent pairing.

Critical weaknesses
  1. The training procedure is simple and preserves single-view inference without introducing additional inference modules. Sharing the base augmentation provides an interpretable way to constrain changes associated with a specific transformation.

  2. The internal controls are useful: the paper separates online synthesis, resampling exposure, and paired consistency, and compares shared-base with independent pairing. The reported four-target mean AUC improves by 3.79 points on GenD and 3.07 points on Effort, supporting the practical usefulness of the combined training recipe within the evaluated setting.

Questions for authors
  1. Does SyRC retain its advantage over the matched baseline on DF40 or a comparable recent facial-forgery benchmark, with training and checkpoint selection still restricted to FF++?

  2. Does shared-base pairing remain better when each pairing construction independently selects its consistency coefficient using the same source-validation rule?

  3. On external target datasets, how do the baseline, exposure-only model, and SyRC compare under unseen resampling kernels and compression conditions?

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

This manuscript proposes SyRC, which mainly combines online SBI with resampling feature consistency. Shared-base pairs focus the consistency constraint on resampling-induced changes. Experiments show improved cross-dataset detection with GenD and Effort while preserving single-view inference.

Critical strengths

The proposed method is simple, readily incorporated into existing detectors, and adds no inference cost.

Critical weaknesses

1. Limited methodological novelty.

The contribution appears incremental over SBI and established consistency regularization. CORE [R1], not cited, already applies the same squared-cosine loss to normalized features of augmented views for face forgery detection (Secs. 2.2–2.3; Eqs. 3, 5). SBI already generates blends online [R2]; the refresh-versus-replay comparison (Sec. 4.3, Table 3) quantifies regeneration within this pipeline but does not make online refresh a new contribution. The remaining differences are shared-base pairing, the replacement schedule, and single-view supervision. Table 5 favors shared-base over independent pairing, but both use the same consistency weight despite different pair discrepancies (Table 5, bottom); separately tuned weights and CORE’s two-view supervision are not evaluated.

2. Limited evidence for the explanation and the practical gains.

The resampling diagnostics are source-domain only; whether the stability gains carry over to target domains and explain the cross-dataset improvements is unclear (Fig. 3, Table 16). With Gaussian blur as the paired transformation, performance is close to training without any paired view, and what makes a transformation effective is not established.

References

[R1] Y. Ni, D. Meng, C. Yu, C. Quan, D. Ren, and Y. Zhao. CORE: Consistent Representation Learning for Face Forgery Detection. CVPR Workshops, 2022.

[R2] https://github.com/mapooon/SelfBlendedImag

Questions for authors
  1. Under matched GenD/SBI settings, does the shared-base advantage over independent pairing persist with separately tuned consistency weights, and how does CORE’s original two-view supervision compare?
  2. Do the stability gains persist on target domains, and what explains DR-128’s advantage over Gaussian blur as the paired transformation?
Reviewer_BaM3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes SyRC, a training recipe for CLIP-based deepfake detectors. It combines online regeneration of self-blended pseudo-fakes with a feature-consistency loss between each augmented image and its downsampled-and-restored version, and leaves inference unchanged. Trained on FF++, SyRC raises GenD’s mean video-level AUC on CDF-v1/v2, DFDCP and DFDC from 89.80% to 93.59%, with a similar gain for Effort.

Critical strengths
  1. The paper frames forgery diversity and stability under post-formation resampling as complementary needs, and addresses both within a single training recipe.

  2. Deriving the resampled view from the already augmented image is a clean design choice. The consistency loss then targets the resampling change alone, rather than also absorbing a second augmentation draw.

  3. The method is training-only. It adds no parameters, keeps single-view inference, and applies to more than one detector.

Critical weaknesses
  1. The contribution is thin. SyRC combines self-blended pseudo-fakes, resizing augmentation and two-view feature consistency, all of which are established. Two-view representation consistency has already been applied to face forgery detection, for example in CORE, which penalizes the distance between representations of two augmented views of the same image; this work is not cited. Most of the gain comes from simply adding SBI to GenD, while consistency and shared-base pairing each add less than one point. The reproduced GenD baseline is also well below the published GenD results, and SyRC lands close to those published numbers, so the improvement over the base method as published may be small.

  2. The evaluation is outdated and favors the method. It relies solely on the classic FF++ protocol with four face-swap datasets from 2019–2020. This is narrower than the base method’s own evaluation, which covered 14 datasets released between 2019 and 2025. CDF-v2 also appears saturated: several methods in the paper’s own Table 2 already reach 95–96% AUC, and SyRC is only 0.16 points above the published GenD result. These datasets are where blending-based training is known to generalize well, because their fakes contain blending artifacts. SBI accounts for most of SyRC’s gain, yet recent and non-blending forgeries are not tested. The stated motivation, robustness to downsampling and restoration, is never evaluated on degraded target data.

  3. The presentation obscures a simple method. The paper introduces many idiosyncratic and inconsistently used terms (e.g., “online refresh” vs. “online formation”). Key claims are stated without magnitude, and the prose is heavily hedged and repetitive. Section 3.4 and Appendix H spend substantial space on an elementary bound that does not explain the cross-dataset gains.

Taken together, addressing these issues would require repositioning against prior consistency-based methods, new evaluations on recent forgery benchmarks, and a substantial rewrite. This goes beyond what can be resolved in the rebuttal period.

(CORE: COnsistent REpresentation Learning for Face Forgery Detection) (DF40: Toward Next-Generation Deepfake Detection)

Questions for authors
  1. How does SyRC differ from CORE and other two-view consistency methods for face forgery detection, beyond deriving the second view from the same augmented image?

  2. Why is the reproduced GenD baseline several points below the published GenD results? What does SyRC add on top of GenD trained with its official setup?

  3. How does SyRC perform on recent and more diverse forgeries, such as the 14-dataset suite used by GenD or DF40? Does the gain hold for forgery types without blending artifacts?

  4. Given that the motivation is robustness to downsampling and restoration, how does SyRC perform on degraded or in-the-wild target data?

  5. Why should consistency under a single DR-128 operation improve cross-dataset generalization, and why does Gaussian blur perform worse? Is there an explanation beyond empirical selection on FF++ validation?

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

Related papers

Discussion (0)

Sign in to comment.