acceptodds
Withdrawn from ICLR 2027

CounterFake: Towards Generalizable MLLM-based Deepfake Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deepfake detection remains challenging under distribution shifts caused by unseen data domains and evolving forgery techniques. Multimodal large language model (MLLM)-based detectors show promising generalization but remain vulnerable to these shifts. We introduce CounterFake, an MLLM-based deepfake detector designed to improve robustness to various domain and forgery shifts. To improve detection of fake images from unseen domains, we develop two complementary counterfactual preference losses: an image-level loss using hard-negative images and a text-level loss using counterfactual rationales. These losses are supported by ReForensics (Reference-guided Forensics Annotation), a comparative captioning pipeline that uses reference images to generate forensic rationales grounded in decisive forgery cues. We further highlight two simple yet important design choices for MLLM-based deepfake detection: (i) weight-space interpolation improves robustness on real images from unseen domains, while (ii) intermediate representation probing reads out authenticity information directly from intermediate language model features while retaining the ability to generate forensic rationales. Extensive experiments on DF40-based benchmarks demonstrate that CounterFake outperforms existing state-of-the-art MLLM-based deepfake detectors, e.g., by 9.1 and 7.9 percentage points in average macro-F1 under cross-domain shifts and joint domain/forgery shifts, respectively.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Deepfake Detection, Media Forensics, MLLMs

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aKVzOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces CounterFake, an MLLM-based deepfake detector designed to generalize across unseen image domains and forgery methods on both real and fake images. It combines reference-guided forensic annotation (ReForensics), textual and visual counterfactual preference optimization, weight-space interpolation, and intermediate-layer linear probing. Extensive experiments demonstrate the effectiveness and superiority of the proposed method compared with existing methods.

Critical strengths
  1. Well-motivated research problem. The paper explicitly addresses distribution shifts affecting both real and fake images, which should be more severe for real image detection I believe. This is a critical issue in the deepfake detection area.
  2. Well-motivated optimization method. Reference-guided annotation and complementary textual/visual preference objectives offer a coherent approach to aligning forensic descriptions with visual evidence for both real and fake images.
  3. Extensive experiments with recent baselines demonstrate the effectiveness of the proposed method.
Critical weaknesses
  1. Methodological novelty. Although I acknowledge the effectiveness and contribution of the designed modules, such as ReForensics, counterfactual preference losses, weight-space interpolation, and intermediate representation probing, they make contributions like combinations of established techniques. The authors should justify their contributions and explain how each module contributes to their contributions.
  2. Potential imbalance during optimization. I acknowledge that the ReForensics and counterfactual preference losses should benefit the detection for both real and fake images, but I doubt there may be potential imbalance introduced during optimization, since by constructing counterfactual pairs, the negative pairs can be constructed from many different combinations, such as many unpaired fake clues, which means the negative space should be larger and have more directions in latent space compared to positive pairs. How did the authors address this issue during optimization?
  3. Post hoc design. I believe the introduction of weight-space interpolation and intermediate representation probing can improve detection performance since there are already many support references. However, I cannot understand why the authors specifically introduce these modules after fine-tuning. It feels like 'extra parts' outside the main claim of this paper. More explanations are needed.
  4. Insufficient attribution of performance gains. As stated above, there are many technical modules introduced. In Table 3, preference optimization improves macro-F1 over SFT by 3.74%, whereas removing interpolation or probing reduces performance by 7.03% and 8.86%, respectively. Does this mean the post hoc designs are more important than the main claim 'counterfake'? More justifications are needed.
Questions for authors

Please see the weaknesses above. My major concerns focus on the method novelty, technical design, and performance attribution. Thus, I currently lean towards a negative rating. Addressing these issues should make this paper clearer and more convincing. I hope the authors address these issues during rebuttal.

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

The paper presents an MLLM-based deepfake detection approach designed for distribution shift. It combines forensic-style data annotation, cross-modal counterfactual preference learning, and lightweight post-processing to generalize evenly on both real and fake samples, while still producing interpretable rationales. However, the training data is still constrained by the demographic and environmental diversity of existing datasets such as FaceForensics++. ReForensics also depends on having a suitable reference image, so it is of limited use for forged images found in the wild whose source is unknown.

Critical strengths
  • A novel, data-backed diagnosis. Instead of vaguely citing "poor generalization," the paper splits distribution shift into real-side and fake-side shifts. It shows that existing MLLM detectors have a severe real/fake recall asymmetry (Figure 7): boosting forgery detection comes at the cost of real-image recall. This leads directly to a complementary design, with a preference loss adding +6.8 pp fake recall and weight interpolation adding +15.4 pp real recall, so motivation and method form a closed loop rather than a collection of tricks.

  • Original cross-modal counterfactual negatives. On the text side, only visual attributes and the conclusion label are changed, while sentence structure and style are kept, which blocks shortcuts based on writing style. On the image side, CLIP retrieves the top-100 visually similar images of the opposite class, and a pretrained MLLM's log p(y_w|I) selects the most plausible-looking one as the hard negative. Each channel fixes one modality and perturbs the other, so they are orthogonal and complementary. Ablations (Table 3a) show that each alone beats plain SFT and that combining them works best.

  • A solid, honest evaluation. Cross-domain testing uses the full Celeb-DF set (92K samples, 32 forgery methods) rather than a hand-picked subset. Cross-domain plus cross-forgery testing uses 7 unseen methods and 4 unseen real-image sources. The paper reports macro-F1 as well as accuracy, and the appendix explains that this exposes single-class collapse: many baselines have decent accuracy but macro-F1 near 0.33, meaning they predict one class for everything. Appendix D gives per-method results and does not hide weak spots.

Critical weaknesses
  1. The largest gains come from two generic, borrowed tricks, not the core innovation. Table 3b shows accuracy dropping from 0.8843 to 0.8209 (−6.3 pp) without weight interpolation and to 0.8025 (−8.2 pp) without the mid-layer probe. Counterfactual preference optimization improves only modestly over plain SFT (0.8479). Weight interpolation comes from Wortsman et al. (2022), and linear probing is standard, and neither is specific to deepfakes. The paper centers its narrative on counterfactual preference learning, but post-processing carries most of the SOTA result, so contribution and emphasis are mismatched and novelty is in doubt.

  2. A key control is missing: the post-hoc tricks were not applied to the baselines. Since interpolation and probing contribute 6–8 pp, the paper should show whether FakeVLM or VERITAS can match it with the same two tricks. Appendix E compares only five vision-only detectors trained on the same data, with no post-hoc-enhanced version of the strongest MLLM baselines. Tables 1 and 2 therefore compare "CounterFake plus two post-processing steps" against the baselines' raw outputs, which is not an equivalent comparison, and the headline 9.1 / 7.9 pp gains may be inflated.

  3. Model selection overlaps with the cross-domain test by method name. Appendix C.1 admits that the validation methods MCNet, SadTalker, and Wav2Lip also appear among the 32 cross-domain subsets. The real-image sources differ (FF++ for validation, Celeb-DF for testing), but the probe layer (11), ridge λ (10³), and interpolation α were all chosen on a set sharing forgery-method names with the test set. This weakens the "unseen forgery" claim, and there is no sensitivity analysis with a fully disjoint validation set.

  4. Probe-injected labels make the rationales' faithfulness doubtful, which conflicts with the interpretability claim. At inference, a linear probe predicts the real/fake label, the label is forced into the label slot, and the model then writes the rationale. The rationale is thus post-hoc rationalization of a decided conclusion, not a natural product of visual reasoning. No causal faithfulness test is run (e.g., flipping the probe label to see whether the model fluently invents opposite evidence), yet Figure 3 self-scores a "faithfulness" dimension, so its validity is questionable. The human study is also small (32 raters × 20 images), uses only samples all four models judged correctly, and has no significance tests or confidence intervals.

  5. ReForensics has limited applicability and reproducibility. It assumes a suitable reference image (a paired real source or a fake from the same method), which is exactly what in-the-wild deepfakes lack. The authors acknowledge this, but their fix (fabricating and annotating fakes from in-the-wild real images) reverts to controlled data and is not validated on real wild distributions. The pipeline relies on Gemini 2.5 Flash and GPT-4o-mini plus manual filtering, and the paper promises no release of code, data, or the 59-class artifact taxonomy, so the annotation pipeline cannot be reproduced. Evaluation stays within the DF40 ecosystem, with no independent benchmark of social-media compression or adversarial perturbation.

Questions for authors

N/A

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

This manuscript proposes CounterFake, an MLLM-based framework designed to improve the generalizability and robustness of deepfake detection under domain and forgery shifts. By introducing an image-level loss based on hard-negative images and a text-level loss based on counterfactual rationales, the proposed method aims to enhance detection generalizability. Furthermore, two components, weight-space interpolation and intermediate representation probing, are introduced to improve the model's robustness to real images and generate more reliable forensic rationales. Extensive experiments under various settings demonstrate the effectiveness and generalizability of the proposed method.

Critical strengths

ReForensics introduces a novel approach to enhancing MLLM-based annotations through reference-image comparisons and a taxonomy of forensic artifacts. The designed loss objectives are well motivated and effective. Furthermore, the proposed weight-space interpolation and intermediate representation probing components effectively improve detection generalizability and the accuracy of forensic explanations.

The proposed method achieves sota deepfake detection performance across various experimental settings. Extensive ablation studies further validate the effectiveness of the selected hyperparameters and designed components. Moreover, the qualitative results demonstrate the promising explainability of the proposed method.

Critical weaknesses

It is suggested that the authors incorporate additional detection and explainability metrics to provide a more comprehensive evaluation and better interpretation of the experimental results.

Questions for authors

Please see the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.