acceptodds
Under review as a conference paper at ICLR 2027

Towards Generalizable Deepfake Detection via Real Distribution Bias Correction

Abstract

Deepfake detection can benefit from the real-image knowledge encoded by pretrained foundation models. Existing approaches seek to retain this knowledge by preserving a subset of pretrained parameters while updating the remaining parameters for deepfake detection. However, freezing pretrained weights cannot freeze their output responses, which can cause channel-parameter mismatch in downstream layers and undermine the use of pretrained knowledge. In principle, gradient-based optimization may mitigate this mismatch, but our analysis reveals that classification supervision overemphasizes a small set of head channels while leaving tail channels weakly constrained. Consequently, the misaligned tail responses can interfere with predictions in unseen domains and hinder generalization. To address this problem, we propose the Real Distribution Bias Correction (RDBC) framework, which uses real-image statistics as reference signals to calibrate the under-constrained channel-response space. Specifically, the Real-Kurtosis-Guided Channel Alignment (RKA) module corrects the accumulated distributional drift at the final normalization layer. Using the input-side kurtosis of real samples as a reference, RKA aligns the output cross-channel response structure and constrains weakly supervised tail-channel responses. The Real-Covariance-Guided Channel Enhancement (RCE) module uses real-feature covariance to propagate the centered deviations of highly responsive channels to other co-varying channels, thereby reducing the reliance on isolated head-channel responses at shallow layers. RCE further samples its statistics from a Gaussian-fitted statistical neighborhood, expanding the coverage of the real-feature reference. Through the complementary calibration of shallow and deep channel responses, RDBC corrects the channel-parameter mismatch while enhancing real-fake separability. Extensive experiments demonstrate that RDBC achieves state-of-the-art performance in cross-domain deepfake detection.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Deepfake Detection, Real Distribution, Generalization, channel-parameter mismatch

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3jPfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Real Distribution Bias Correction (RDBC), a framework for generalisable deepfake detection built on Effort (CLIP ViT-L/14). A key finding is that frozen pretrained weights do not freeze their responses, allowing weakly constrained tail channels to work against the correct class on unknown domains. RDBC calibrates the channel space using statistics from real images via Real-Kurtosis-Guided Channel Alignment and Real-Covariance-Guided Channel Enhancement. After training on FF++, the model is tested on six datasets along with ablation studies, robustness tests, and a drift analysis.

Critical strengths
  1. The paper makes a useful observation that frozen weights do not imply frozen output responses. This addresses parameter-preserving fine-tuning methods, which remains underexplored. The channel-wise analysis in Fig. 1 translates this into a testable hypothesis. This is the main contribution of the paper, since the perspective could be relevant beyond deepfake detection.

  2. RDBC requires no target-domain data and adds little overhead, but still improves over Effort in all reported settings, in some settings by a large margin (e.g. 6.62% AUC on DFDC at video level). The ablation in Table 5 shows that both modules contribute individually, which supports the practical value of the method.

Critical weaknesses
  1. The main results are not comparable to the state of the art. The central claim relies on a reimplementation of Effort that falls short compared to the original Effort paper, by almost 10 AUC points on DFDC at video level. RDBC stays below the published Effort results at nearly all points. It is therefore hard to say whether RDBC actually improves on the state of the art. Section 4 states that the same protocol as the backbone is used, but the reported setup differs, which also makes the values taken from other papers hard to compare. The origin of the numbers is sometimes unclear and is even mixed within rows. Finally, all results come from a single seed without variance, while some margins are below one point. This is critical, because state of the art cross-domain performance is the main claim of the paper.

  2. It is not shown that the gains actually come from correcting tail-channel mismatch. The mismatch is never clearly defined, and the tail improvements in Table 7 are tiny, even smaller than in the head channels, and measured on one domain without error bars. The arguments for the modules are also weak. LayerNorm does not change kurtosis, so RKA only reacts to the frozen scaling parameters, and kurtosis is driven by extreme channels rather than the tail. The separability claim for RCE holds for any rescaling of the features. Without further evidence, what remains are two regularizers that help Effort, which is much less than the paper acually claims.

Questions for authors

Clarification:

  1. Why does your reimplementation of Effort fall below the published results, by almost 10 AUC points on DFDC at video level? Could you report RDBC on top of Effort under the original protocol?
  2. Please clarify for every entry in Table 2 and 3 whether it was reimplemented or copied from another paper.
  3. Could you report mean and standard deviation over multiple seeds for the main results and the ablation in Table 5?
  4. What are the pre- and post-LayerNorm kurtosis values for the unmodified pretrained CLIP? Since LayerNorm itself does not change kurtosis, what does RKA constrain, and how does it affect the tail channels specifically?
  5. How does RDBC compare to MixStyle and simple feature distillation toward frozen CLIP on the same backbone and protocol?

Suggestions for future work:

  1. Evaluation on general AI-generated image benchmarks such as GenImage or UniversalFakeDetect.
Reviewer_F69COpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Real Distribution Bias Correction (RDBC) to address channel-response drift when adapting pretrained foundation models for face forgery detection. The authors argue that freezing pretrained parameters does not necessarily preserve feature responses, while classification supervision primarily concentrates on a few dominant channels, leaving less-constrained channels potentially susceptible to interference on unseen domains. RDBC introduces Real-Kurtosis-Guided Channel Alignment (RKA) to align the cross-channel kurtosis of real-sample features before and after the final frozen post-LN layer. It also incorporates Real-Covariance-Guided Channel Enhancement (RCE), which transforms high-response patches in shallow layers using sampled real-feature statistics during training. Experiments show improvements over Effort across multiple cross-dataset evaluations.

Critical strengths
  • The motivation is clear and addresses a common practical challenge in generative replay for face forgery detection.

  • The experiments are comprehensive, with ablation studies providing evidence for the contribution of each component.

  • Under the assumptions considered in the paper, the proposed loss appears effective in reducing the influence of generator-induced artifacts on the detector.

Critical weaknesses
  • This paper uses the kurtosis of the current input features as the alignment reference, rather than statistics from the original CLIP representations. Although detaching this reference prevents gradients from propagating through it, the reference itself can still change during training. Further explanation or analysis is needed to establish how this evolving target restores compatibility with pretrained representations.

  • It holds the classifier fixed to isolate the effect of response changes, whereas Figure 1 compares separately trained models using their respective channel-contribution rankings. Classifier weights, channel selection, and feature responses may therefore vary simultaneously. The current analysis does not sufficiently distinguish the correction of harmful response mismatch from general regularization or classifier reweighting.

  • The paper does not clearly specify test-frame sampling, video-score aggregation, checkpoint selection, or the exact insertion point of RCE. On DFDC, the improvement is 1.83% at the frame level but 6.62% at the video level. Clarifying the aggregation protocol is particularly important for interpreting this difference and reproducing the results.

  • The DFDC result at (n=128) in Figure 3 appears to be approximately 83%, whereas Table 5 reports 81.11% for the complete RDBC model. Since the text states that (n=128) is used in subsequent experiments, the authors should clarify whether these results were obtained under different configurations.

  • Table 6 reports accuracy only on real samples without specifying the classification threshold or its selection procedure. This metric measures the correct acceptance rate of real samples at a particular operating point. Without the corresponding fake-sample recall at the same threshold, the improvement cannot readily be attributed to better modeling of the real distribution. The threshold, prediction unit, and data used for threshold selection should be specified, together with the corresponding fake-sample detection results.

Questions for authors

see weakness

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

This paper proposes Real Distribution Bias Correction (RDBC) for cross-domain deepfake detection using a partially frozen pretrained visual encoder. It attributes generalization difficulties to channel-response drift and insufficient supervision of low-contribution channels. RDBC combines Real-Kurtosis-Guided Channel Alignment (RKA), which matches real-sample kurtosis across the final frozen normalization layer, with Real-Covariance-Guided Channel Enhancement (RCE), which transforms selected shallow patch features using sampled real-image statistics. Experiments assess frame-level and video-level detection, robustness to image degradations, and the contributions of individual modules.

Critical strengths
  1. The paper addresses an important generalization problem and offers a concrete channel-level perspective on adapting pretrained visual representations for deepfake detection.

  2. The evaluation covers multiple target datasets, both frame-level and video-level metrics, image corruptions, and module ablations.

  3. The statistical reference uses only source-domain real images, and inference employs deterministic statistics without requiring target-domain data.

Critical weaknesses
  1. The evidence does not establish that channel-parameter mismatch causes poor generalization. Equations (1)–(6) describe response changes under fixed mappings, but do not distinguish harmful drift from useful adaptation. Figure 1 and Table 7 provide correlational evidence. Please add targeted interventions, such as unfreezing only post-LN affine parameters or restoring selected pretrained tail responses, and measure their effects on target-domain AUC and classification margins.

  2. RKA matches a single kurtosis statistic in a 1024-dimensional representation. Matching kurtosis does not establish channel-wise alignment, since different channel arrangements can share the same statistic. Moreover, involving every channel does not demonstrate effective tail-channel supervision. Please report head/tail gradient magnitudes from RKA separately from cross-entropy and quantify channel-wise response drift under RKA alone.

  3. RCE transforms centered features as v^=(I+C)v\hat v=(I+C)v, where v=x−μv=x-\mu, which can amplify dominant directions without balancing channel responses. For example, C=diag⁡(100,1)C=\operatorname{diag}(100,1) maps v=(1,1)⊤v=(1,1)^\top to (101,2)⊤(101,2)^\top, increasing concentration. Please measure response concentration before and after RCE and compare full, diagonal, and normalized covariance variants. The dependence of enhancement strength on feature scale also requires justification.

  4. The current ablations do not establish an advantage over existing feature-statistics augmentation. Comparing no enhancement, fixed statistics, and sampled statistics cannot isolate covariance coupling from generic stochastic regularization. Please compare closely related statistical augmentation approaches using the same backbone, insertion point, and training budget, and include direct resampling of empirical subset-statistic pairs to assess whether Gaussian fitting provides additional benefit.

  5. The evaluation protocol is insufficiently specified to interpret the larger video-level improvement on DFDC: 6.62 percentage points versus 1.83 at frame level. Please specify dataset versions, test subsets, frame counts, sampling, video aggregation, and whether both metrics use identical videos. Also explain source-only checkpoint selection and identify protocol differences among results collected from original papers, third-party reproductions, and the authors’ implementations.

  6. Only one random seed is reported, despite additional randomness from subset selection and statistical sampling. Several gains are small, including 0.23 percentage points on WDF and a Tail-10% drift reduction from 0.2133 to 0.2124. Please report multiple independent training runs for the baseline, full method, and key ablations, together with confidence intervals for AUC differences using video-level resampling.

  7. Real-only accuracy measures real-class recall and can improve by favoring real predictions while reducing fake detection. Table 6 specifies neither the threshold nor evaluation granularity and omits corresponding fake-class performance. On Celeb-DF, 39.62% real accuracy implies a 60.38% false-positive rate. Please report both class recalls, balanced accuracy, and target-domain performance at thresholds determined exclusively on source validation data.

  8. The real-only design is central to the motivation but is not isolated by the module ablations. The improvements could arise from generic feature perturbation or regularization rather than the special value of real-image statistics. Please compare real-only, fake-only, and mixed statistical references with matched sample counts and transformation strengths, and compare RKA applied only to real samples against alignment applied to both classes.

Questions for authors
  1. What existing evidence distinguishes correction of harmful tail-channel mismatch from generic regularization or feature rescaling?
  2. Were frame-level and video-level AUCs computed using identical test videos, and how were frame predictions aggregated and checkpoints selected?
  3. What threshold was used for Table 6, and what fake-class recall accompanies the reported real-class accuracy?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.