DiffusionFake: Enhancing Generalization in Deepfake Detection via Guided Stable Diffusion

Summary

This paper introduces DiffusionFake to address the challenge of generalization in face forgery detection by revisiting the generative process of deepfakes. DiffusionFake reverses the generative process by injecting features into a pre-trained Stable Diffusion model to reconstruct source and target images. The plug-and-play framework integrates with existing detectors, improving cross-domain generalization without adding inference parameters. Experiments demonstrate significant performance improvements across various architectures.

Strengths

1. This paper presents a novel approach by designing a method from the deepfake creation perspective, introducing a reverse method, and utilizing pre-trained Stable Diffusion knowledge to address information loss during reconstruction. It effectively extracts features using a guide module to obtain source-related and target-related information through reconstruction, which is logically sound. 2. The paper proposes a plug-and-play functionality that enhances the generalization ability of multiple detection networks without adding parameters, facilitating deployment. 3. This paper is well-written. The authors offer an easy-to-follow presentation with a well-structured format, while important visualizations and figures are clearly provided.

Weaknesses

1. The training set seemingly contains instances where the source and target ground truth are identical for the same image. The paper should address how this trade-off is handled. 2. The discussion on the weight module could be clearer and more detailed. 3. In the visualization figures (Fig. 3), the reconstructed source images appear to differ from the ground truth (even the training sample). The authors should explain the reason for this discrepancy. 4. More ablation studies should be conducted. For example, the structure of the feature-transform module is insufficiently evaluated.

Questions

1. Would the performance improve and convergence speed up if the parameters of Stable Diffusion were unfrozen during training? 2. How significant is the impact of Stable Diffusion's pre-trained knowledge on the overall method? Have other pre-trained diffusion methods been tested? For more questions please refer to the Weaknesses.

Rating

7

Confidence

5

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors have adequately addressed the limitations.

Summary

This paper proposes DiffusionFake which can harnesses the power of Stable Diffusion to guide the forgery detector in learning disentangled source and target features inherent in Deepfakes. The features of the detection networks are processed through the target and source transformation modules, and then injected into the Stable Diffusion to reconstruct the source and target images. Through the proposed strategy, DiffusionFake can enhance its ability to handle unseen forgeries without compromising efficiency.

Strengths

The idea of using SD-based reconstruction to enhance detection accuracy is rational.

Weaknesses

1. The weighting module is supervised by the similarity between z and $z_s$/$z_t$. However, the latent representation of stable diffusion is not linear. I think the assumption of this constraint is not always correct. 2. There is no analysis for the influence of the loss weight in Eq. 13 3. There is no cross-model evaluation, i.e., the samples during testing are generated by a method that has not been seen during the training. 4. The evaluation is only conducted on the images with low resolution.

Questions

1. How to decide the optimal loss weights in Eq. 13? 2. How is the performance if we conduct cross-model evaluation? 3. How is the performance when the input image has high resolution?

Rating

5

Confidence

5

Soundness

2

Presentation

2

Contribution

2

Limitations

If some attackers have known the training details of the detector and use it as the metric to enhance the performance of deepfake, it will be harder to detect them.

Summary

This paper adopts a novel plug-and-play framework that reverses the generative process of face forgeries to enhance the generalization of detection models. Extensive experimental results from several datasets demonstrate that this method has achieved very competitive performance.

Strengths

- The motivation of this paper is clear and the approach to implementing it is straightforward and understandable. - This paper attempts to introduce Stable Diffusion into deepfake detection, unifying the generation model and the detection model end-to-end, which is very interesting and innovative, and provides a new perspective for this field. - The overall structure of this paper is relatively reasonable and the context of the paper is relatively clear.

Weaknesses

- Some recently published papers about enhancing generalization in deepfake detection are not cited and discussed, such as: [1] Dong, Shichao, et al. "Implicit identity leakage: The stumbling block to improving deepfake detection generalization." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023. [2] Yu, Bingyao, et al. "Discrepancy-aware meta-learning for zero-shot face manipulation detection." IEEE Transactions on Image Processing (2023). [3] Chen, Liang, et al. "Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022. - Many of the experimental results in Table 1 provided by the authors in the paper are inconsistent with or may be lower than the experimental results in the original paper. It is best for the authors to provide detailed, clear and convincing explanations. - In order to better understand the performance of the proposed method, it would be best if the authors could provide sample visualization results of misprediction and give corresponding theoretical analysis. - Please unify the format of references. At least make sure that the reference formats of conferences and journals are consistent. - If the author can answer and revise the relevant questions in the final version, I will consider increasing the final score in the next round.

Questions

Please standardize the capitalization of English letters in the references. Many abbreviations of proper nouns are incorrect, such as Aunet (AUNet).

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have adequately addressed the limitations and potential negative societal impact of their work.

Summary

The paper investigates the task of deepfake (especially, face swapping) identification. It proposes to utilize current image generation model to reconstruct source and target profiles from embedded features. The authors did thorough experiments and prove quantitatively that the proposed DiffusionFake method outperforms previous baselines.

Strengths

- The problem definition is clear and well-motivated. The proposed solution is also quite simple and intuitive. - All the figures and diagrams are well-designed. Specifically, the first few figures explain the pipeline clearly. Figure 4 and 5 visualize the results in a nice way. - Quantitative results show that the proposed method significantly outperform baselines.

Weaknesses

- Field of application is limited to face swapping with two identities, while claiming a "DeepFake detector". - The Related Work section is not very clear in connecting the current work with previous ones. See questions. - Writing quality is in general good. It might benefit from omitting some details (how many layers?) and provide more intuitive explanation (why this design works?) - Experiments are thorough, but only on two detection metrics. It would be great to instead see something measuring reconstruction quality.

Questions

- Why DFD dataset's best performance happens with a different model architecture than all others? Is there anything special about this dataset and its performance? - In Figure 3A, even with training samples, we can see that the expression is not quite well reconstructed. The target GT expression (mouth slightly open) is "recovered" as source expression. Is this a common observation? How well does the model preserve expression? - Using AUC and EER as metrics are good (and classical) regarding this task. However, as in the ablation study, it would be interesting to see some numerical metrics measuring the reconstruction quality. Is this quality highly-correlated with the detection accuracy? - Is this the first ever work that takes advantage of generative model for face forgery detection? What are some of the most important baselines? Qualitatively how does the current methods outperform the baselines? (i.e. what are some characteristics of those challenging cases that baselines cannot do but this approach can?)

Rating

7

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

The authors already mentioned a list of limitations at the end of their paper. - Not applicable to blending of multiple identities (>2) or partial manipulation. - Require paired images, not self-supervised. - Could be used as a discriminator to develop harder to distinguish deepfake images. And also two more: - The dataset they use might have limited diversity -- looks like all the examples in the paper are white. A bit concerned about the generalizability to all ethical groups, genders, ages, etc. - The word "Deepfake" might have a slightly broader meaning than face swapping these days. This paper is only detecting face swapping, not "generated faces" in general. It might worth being a bit conservation in claiming the contribution.