FreqBlender: Enhancing DeepFake Detection by Blending Frequency Knowledge

Summary

The generalization of DeepFake detection can be addressed by enhancing training data with synthetic fake faces, known as pseudo-fakes. Traditional methods generate these faces by spatial blending operations. However, the limitations of these methods are that they ignore to simulate the frequency distribution, where additional important forgery clues might be found. This paper introduces an interesting method called FreqBlender, which attempts to blend proper frequency knowledge to enhance the effectiveness of pseudo-fake faces. This method involves identifying forgery clues in the frequency domain and blending the corresponding frequency components from fake faces into real faces. This process is challenging due to the variability and spread of forgery clues across different frequency ranges. To achieve frequency decomposition, the authors propose a Frequency Parsing Network (FPNet) that can adaptively partition the frequency domain. The network, consisting of an encoder and three decoders, extracts semantic, structural, and noise information from the faces. Training FPNet is difficult due to the lack of ground truth for frequency distribution, so the authors devise a novel training strategy leveraging inner correlations among different frequency components. Once trained, FPNet can extract the structural information of a fake face's frequency component and blend it with a real face to generate a pseudo-fake face. This method complements existing spatial blending methods and improves detection performance on multiple DeepFake datasets, demonstrating its effectiveness.

Strengths

1. This papeer enhancing the generalization of DeepFake detection is a critical and current issue in AI safety, as real-world DeepFakes are likely generated by unknown models. This method addresses this challenge by taking an innovative approach, creating pseudo-fake faces through blending frequency knowledge rather than the conventional spatial knowledge used in existing methods. 2. This work identifies the limitations of existing methods, noting that spatial blending only mimics the visual similarity of wild fake faces but overlooks their frequency characteristics. By analyzing the frequency distribution, the authors describe a Frequency Parsing Network (FPNet) to parse the frequency range carrying forgery traces. Interestingly, this method doesn't require wild fake faces for blending; instead, it uses existing spatial-based pseudo-fake faces as surrogates. 6. The method's efficacy is validated across multiple recent DeepFake datasets, demonstrating its effectiveness with various backbones and complementing existing spatial-blending methods. This showcases its robustness and practical applicability in enhancing detection performance.

Weaknesses

This paper proposes an interesting and intuitive method that seems reasonable to me. Here are some recommendations for future improvements: 1. Investigate more fine-grained frequency ranges to uncover the detailed composition of artifacts. This could potentially enhance the model's performance by providing a more nuanced understanding of forgery clues. 2. Develop algorithms to enhance the explainability of the model’s decisions. This will help users understand how and why a particular face is identified as a DeepFake, increasing trust and usability in practical applications. 3. Extend the method to handle other types of forgery operations beyond face-swapping, such as whole-face synthesis and attribute editing. This will broaden the method's applicability and ensure it remains effective across a wider range of DeepFake techniques. 4. Utilizing LLMs may help you parse the frequency more precisely, as LLMs can provide the prior knowledge which may compensate for the lack of ground truth when training FPNet.

Questions

1. As described in Section 3, the statistics of the frequency distribution are calculated using an azimuthal average, which includes a logarithmic transformation and the calculation of azimuthally-averaged flux in circular annular apertures. Could you provide more details on why placing the center of the circular annular aperture at the top-left corner of the frequency map results in a one-dimensional spectrum diagram? 2. The FPNet uses three independent decoders to analyze the frequency domain. The output of the decoders is not clearly explained, as it is visualized in grayscale as the frequency map in Fig. 5. I understand that the output is a soft mask ranging from [0,1] for the frequency map. Therefore, I suggest using a color version to replace the DCT map in Fig. 5. 3. In the discussion on the loss of authenticity determination, the authors aim to create two sets of faces, with and without forgery traces. Does the set C_r represent the face sets that do not contain structural information of fake faces, while C_f corresponds to the face sets that have structural information of fake faces? 4. In Prior and Integrity Loss, the final term is designed to ensure that the combination of three frequency components covers the entire frequency domain. This term sums these frequency components and calculates the distance from a mask of 1. However, this term might not guarantee that the sum of these frequency components equals 1, but it can reduce deviations. What is the rationale behind this design??

Rating

6

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

This paper lists its limitations in main body. Since the proposed method is designed to address the limitations of existing spatial-blending techniques. it operates under the assumption that faces are forged using face-swapping methods. It unlikely tackle other types of forgeries, such as whole-face synthesis.

Summary

This work studies generalizable deepfake detection. The proposed method is motivated by a new data augmentation method that blends real and fake faces in the frequency domain. The paper claims the forgery can be found in three different frequency bands and proposes an unsupervised learning method, Frequency Parsing Network. Empirical results indicate that the proposed method's performance achieves results comparable to state-of-the-art.

Strengths

1. unsupervised way of learning the frequency component, which can favor its usage.  2. I like the analysis from Fig 2 and 3, showing the difference between real and fake residing in the high-frequency band. 3. I enjoy the problem formulation as well as the learning object in section 4.2.

Weaknesses

1. the proposed method does not achieve state-of-the-art detection performance in table 1. 2. table 2 only has a few methods on the FF++ dataset. It should not be the case that SBI is only a baseline to compare.  3. the proposed approach might be limited in standard face-swap fake face, or gan-generated face. How about diffusion model generated face? 4. in terms of approach. the proposed method works when capturing semantic, structural and noise information. What if it fails to capture any one of those?

Questions

1. line 31 says the motivation of simulating various blending artifacts. However, these artifacts are quite "old" compared to that from current diffusion models. That says, how would your method perform on the face generated by stable diffusion? such as instantid, photomaker? 2. did you compare with DIRE and HiFi-Net? R1: DIRE for Diffusion-Generated Image Detection, ICCV2023 R2: Hierarchical fine-grained image forgery detection and localization, CVPR2023

Rating

4

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

I am impressed by the learning objective formulation, but I am inclined to reject it because it does not achieve reasonable sota performance and does not show generalization to diffusion-generated images.

Summary

This paper proposes FreqBlender, a new method to generate pseudo-fake faces that effectively simulate the frequency distribution of wild fake faces. Unlike common blending techniques done in the spatial domain, their method blends frequency knowledge. An analysis is conducted, showing that three frequency components are present in faces, namely semantic information, structural information, and noise information. They demonstrated that structural information contains forgery traces. To this end, the first stage of their method employs FPNet, a novel architecture built to disentangle the input fake face into the three different frequency components. As no ground truth exists for this task, carefully crafted objectives provide the necessary supervision. In the second stage, the trained network is used to parse the frequency components, and the structural component is extracted from a given fake face. It is then blended into a real face to obtain the pseudo-fake. The method outperforms the state-of-the-art across different relevant datasets.

Strengths

Originality: This method is the first to mine frequency structural information and propose a way of generating pseudo-fake faces that mimic the frequency distribution of fake faces. Quality: The method is well evaluated on different datasets, and the results are consistent. Clarity: The paper is well written, and the method is well explained. Figures, along with a preliminary analysis of the frequency components present in faces, are provided, which help to understand the method. Significance: This method provides a new pseudo-fake mechanism that complements spatial blending, which is the reference in the literature. The authors have shown that their method improves the results of different frame and spatial blending-based methods such as DSP-FWA, I2G, Face-Xray, and SBI. The authors claim that the code will be released in the future. To ease the reproduction and adoption of this technique, we encourage them to also release the pretrained weights of FPNet.

Weaknesses

References Missing: Some references are missing, for example, methods that are frame-based and only use real faces during training [1, 2], as well as other recent detectors [3, 4]. Additional Overhead: During the training of the detector, the inference of FPNnet (1 encoder and 3 decoders) is required for generating a pseudo-fake. This adds an overhead during training. The authors should include an efficiency analysis. [1] Li et al., Pixel bleach network for detecting face forgery under compression, in IEEE Transactions on Multimedia 2023 [2] Larue et al., SeeABLE: Soft Discrepancies and Bounded Contrastive Learning for Exposing Deepfakes, in ICCV 2023 [3] Dong et al., Implicit Identity Leakage: The Stumbling Block to Improving Deepfake Detection Generalization, in CVPR 2023 [4] Guo et al., AltFreezing for More General Video Face Forgery Detection, in CVPR 2023

Questions

Clarification Needed on FPNet Training: While the authors explain that for deepfake detection, pseudo-fakes generated using SBI are used as a substitute for fake faces from FF++, the authors should include the results when real fake faces from FF++ are used in Table 1. Use of PixelShuffle: Why is a PixelShuffle used in the decoder of FPNet? FPNet Direct Application: Can the FPNet be used directly for deepfake detection? What happens when a real face is input into the FPNet? Ablation Study and Frequency Knowledge: The authors successfully conduct an ablation study on the backbone (Table 5) and show that results improve consistently when compared to SBI. Why does this method introduce frequency knowledge when all the tested backbones are spatial (e.g., EfficientNet-b4)? It would be interesting to compare the proposed method with a built-in frequency-based backbone (e.g., Face Forgery Network (F3-Net) [3], AFFGM [1], or the multi-scale high-frequency feature extractor from [2]). Results on Highly Compressed Data: What are the results of the method on highly compressed data, i.e., FF++ LQ? [1] Li et al., Frequency-aware Discriminative Feature Learning Supervised by Single-Center Loss for Face Forgery Detection, in CVPR 2021 [2] Luo et al., Generalizing Face Forgery Detection with High-frequency Features, in CVPR 2021 [3] Thinking in Frequency: Face Forgery Detection by Mining Frequency-aware Clues, in ECCV 2020

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Acknowledgement of Limitations: The authors acknowledge the usual limitations of pseudo-fake generation methods, i.e., the hypothesis that the test face is crafted using face-swapping techniques may not hold when the test face is generated using other techniques. Discussion on Societal Impact: The race between attackers and defenders is a well-known issue in the field of deepfake detection. The authors should discuss the potential negative societal impact of their work and how it could be exploited by attackers to generate more realistic deepfakes by simply training their generators to fool the proposed method.

Summary

This paper have introduced an effective way to improve the generalization of DeepFake detection via generating pseudo-fake faces by blending frequency knowledge. The proposed approach achieves state-of-the-art (SOTA) results on various deepfake detecion datasets.

Strengths

1) This paper attempts to combine frequency domain information and spatial domain information to deal with deepfake detection, which is very interesting. Spatial domain blending is very common, but it is still very rare to use it in the frequency domain, which is quite innovative. 2) The writing of this paper is easy to understand and the logic is clear. 3) The experiments in this paper are sufficient and effectively support the author's theoretical basis.

Weaknesses

1) The idea of ​​this paper is similar to some already published papers, such as [1], [2] and [3]. I hope this paper can cite and further analyze them: [1] Tan, C., Zhao, Y., Wei, S., Gu, G., Liu, P., & Wei, Y. (2024, March). Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 5, pp. 5052-5060). [2] Yu, B., Li, W., Li, X., Lu, J., & Zhou, J. (2021). Frequency-aware spatiotemporal transformers for video inpainting detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 8188-8197). [3] Frank, J., Eisenhofer, T., Schönherr, L., Fischer, A., Kolossa, D., & Holz, T. (2020, November). Leveraging frequency analysis for deep fake image recognition. In International conference on machine learning (pp. 3247-3258). PMLR. 2) Moreover, Ref.30 is an important and widely cited work that introduces the frequency domain into deepfake detection. I don’t quite understand why the author does not compare and analyze it with this work. 3) The source information, page numbers, publishers, etc. of many references are incomplete, and some even are incorrect. For example, [4] should come from CVPR2024 instead of arXiv. [4] Choi, J., Kim, T., Jeong, Y., Baek, S., & Choi, J. (2024). Exploiting Style Latent Flows for Generalizing Deepfake Video Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 1133-1143). 4) The faces in Figure 6 should have been resized. It would be better if the author could show the original resolution image to help readers better understand the experimental results. 5) Why is the fluctuation of λ3 so much larger than that of other parameters? I hope the author can provide the corresponding theoretical basis and further detailed analysis. 6) 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

Some images in the paper will be a little blurry after zooming in, so it is best to convert all images to pdf or eps format.

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.