Can We Leave Deepfake Data Behind in Training Deepfake Detector?

Summary

This study introduces a novel training strategy for Deepfake detection using real, blendfake, and deepfake datasets. By designing an oriented progressive regularizer and a feature bridging module, the proposed approach effectively extracts forgery information from the training data, resulting in enhanced generalizability.

Strengths

The proposed method categorizes forgery faces into several types: SBI, CBI, and Deepfake faces, each containing distinct forgery artifacts, such as blending clues, identity inconsistencies, and generative fingerprints. The fine-grained learning scheme encourages the model to learn representative features from the training data, thus achieving robust and general face forgery detection.

Weaknesses

1. The method employs a progressive transition from real to blendfake to deepfake samples. However, the necessity of continuity in these features remains unclear. The transition from real to fake faces, as depicted in Fig. 2, appears conceptually weird. The rationale behind the feature bridging and transition design is not well-explained. The progressive transition between adjacent anchors seems unusual, and the reasoning for a continuous rather than discrete transition is not justified. 2. Despite the generative artifacts present in deepfake data, it remains ambiguous why directly incorporating blendfake and deepfake data during training degrades performance. The authors suggest that direct VHT may fail to disentangle the learned representation in the latent space, but no experiments support this claim. 3. Fig. 1(b) does not appear to be an experimental result, which is crucial for validating the work's motivation. 4. In Line 44-45, the authors raised a question “Can the blendfake face entirely substitute the actual AI-generated deepfake face in training deepfake detectors?” However, this question has already been addressed by Face X-ray and SBI, which successfully use blendfake data to train general models. 5. The term A^T in Eq. (6) is not explained. 6. It is unclear why features augmented with noise should be mapped to a more fake distribution. 7. More Deepfake datasets, such as WildDeepfake and DeepForensics-1.0, should be included in cross-dataset evaluations. 8. For robustness evaluations in Table 6, the method should be compared with recent state-of-the-art deepfake detection methods, and more severity levels for each perturbation type should be included to mimic complex real-world scenarios.

Questions

1. In Table 4, is R2B2D the same as D2B2R? 2. What is the real/fake label of the mix of F_r and F_s?

Rating

4

Confidence

5

Soundness

2

Presentation

3

Contribution

2

Limitations

N.A.

Summary

The authors introduced a method aimed at detecting deepfakes. Their approach, known as Oriented Progressive Regularizor (OPR), employs a progressive transition strategy. This strategy is designed to enable the model to effectively train on a combination of blendfake and deepfake data, ultimately leading to improved performance. The experimental results indicated that this method surpasses current state-of-the-art (SOTA) approaches when tested on deepfake datasets.

Strengths

The paper provides a fresh perspective on the well-known problem of deepfake detection, which should be appreciated. The paper is mostly well-written. The arguments and results presented are easy to understand. The authors performed an extensive evaluation.

Weaknesses

Argument on Blendfake: The argument that blendfake data alone is sufficient for training deepfake detectors is based on empirical observations on certain datasets or benchmarks with some particular deepfake detection models and may not hold universally. I would suggest toning down that claim or providing the exact conditions when this argument holds. CDFv1 vs CDFv2: I believe that using CDFv1 for evaluation may not be necessary. It would have been more beneficial to utilize a different deepfake benchmark dataset such as FakeAVCele, DFD from Google/Jigsaw, or RWDF-23 (please refer to this repository for additional information https://github.com/Daisy-Zhang/Awesome-Deepfakes-Detection). The same applies to DFDC and DFDCP. I advocate for incorporating more diversity in the selection of benchmark datasets. In essence, the authors compared against three datasets instead of five, which is still an acceptable number. Datasets: The authors utilized widely known deepfake datasets from 2019 in their research. However, considering the rapid advancements in deepfake technology since then, I believe these datasets may no longer accurately represent the current landscape. It would be valuable for the authors to include an assessment of their method using real-world deepfake videos sourced from social media and other online platforms. By doing so, they can demonstrate the effectiveness of their proposed solution in addressing contemporary and future iterations of deepfakes. Progressive transition: Currently, the Progressive transition goes like this "Real --> Blendfake (SBI) --> Blendfake (CBI) --> Deepfake". I could imagine it being further extended to have addition of compression or adversarial artefacts (i.e., "Real --> Blendfake (SBI) --> Blendfake (CBI) --> Deepfake--> compression and other artefacts"). That way one could really see a generalisable pipeline that could incorporate the variance in the types of deepfakes available on social media and will greatly increase the quality of the work.

Questions

See the above comments.

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

See the above comments.

Summary

This paper investigates the generalization ability of deepfake detectors and proposes a novel training approach using "blendfake" data to enhance the model's learning of generic forgery artifacts. The authors point out that existing state-of-the-art methods do not incorporate deepfake data in their training process, which contradicts previous empirical observations. The paper introduces an "Oriented Progressive Regularizor" (OPR) to establish constraints on anchor distribution and proposes feature bridging to facilitate smooth transitions. Experimental results indicate that the proposed method effectively utilizes forgery information from both blendfake and deepfake.

Strengths

- Proposes a new training method that may enhance the generalization capability of deepfake detectors. - Introduces OPR and feature bridging techniques to improve the model's recognition of forgery features.

Weaknesses

- The attribution of the unorganized latent-space distribution lacks comprehensive experiments. - There are some minor writing issues, such as the consistency of using SOTA and SoTA.

Questions

- In ablation Table 2, the AUC of VHT is lower than that of BF-only in cross-dataset comparison, which is opposite to the statement in line 230. - Are all the training sets for the SOTA model comparisons the same? How do you control the blendfake and deepfake training datasets for different models?

Rating

4

Confidence

3

Soundness

2

Presentation

3

Contribution

2

Limitations

The authors believe that the reason VHT performs worse than blendfake-only is due to the unorganized latent-space distribution. Although the results indicate that the proposed method is effective, there is a lack of detailed experimental validation for attribution.

Summary

The paper explores the utilization of blendfake and pseudo-fake data in training deepfake detectors. It argues that the significance of deepfake samples has been underestimated due to insufficient exploration. To better exploit both pseudo-fake and deepfake data, the paper introduces a progressive transition from "real to blendfake to deepfake" and proposes a hybrid training scheme. This scheme includes an oriented progressive regularizer (OPR) to model the transition and a feature bridging strategy to simulate a continuous transition.The paper explores the utilization of blendfake and pseudo-fake data in training deepfake detectors. It argues that the significance of deepfake samples has been underestimated due to insufficient exploration. To better exploit both pseudo-fake and deepfake data, the paper introduces a progressive transition from "real to blendfake to deepfake" and proposes a hybrid training scheme. This scheme includes an oriented progressive regularizer (OPR) to model the transition and a feature bridging strategy to simulate a continuous transition.

Strengths

1.The paper is well-motivated, and the proposed solution is both intuitive and effective. 2.The experiments robustly demonstrate the rationality and effectiveness of the proposed design.

Weaknesses

1. Choice of Blend Algorithms: The paper does not provide sufficient explanation or discussion on the choice of blendfake image algorithms (SBI and CBI). As mentioned in Section 2.2, there are many other methods for crafting blendfake images. Would these methods be effective as well? 2. Interpolation Strategy: In Section 3.2, the paper introduces an interpolation strategy to achieve a smoother transition from real to deepfake. Why was interpolation performed at the feature level, and would setting multiple mixing parameters (alpha) for more interpolations further improve performance? 3. Possible Typos: There might be a typo on line 149, $M_a$.

Questions

See in weakness

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

See in weakness