1. Introduction

Figure 1. The first row shows the common CV dataset, the second row shows the human facial expression dataset, and the third row shows the deepfake dataset.

Measuring the relevance of each deepfake dataset is crucial. To address this problem, We establish the KCE-System. It uses the Xception model ^[5] as a forgery feature extractor that maps various deepfake images into the feature space. Then, we use PCA for dimensionality reduction and the K-means method for clustering. We use these clustered datasets to retrain the Xception model and use the Calinski Harabasz Index ^[6] to judge the models' performance. To improve the credibility of the experimental results, we repeat part of the experiments on The Frequency in the Face Forgery Network (F3-Net) ^[7] and Residual Neural Network (ResNet) ^[8]. We also combine these deepfake datasets based on forgery method labels as a control group.

Our experiments prove that some forgery category labels of the same name differ significantly across different datasets. When the forgery method of the deepfake dataset is unknown, the KCE-System can achieve better generalization performance by training on merged datasets based on closer feature distances.

2. Deepfake Datasets

Table 1. The standard deepfake datasets. The symbol * represents the number of pictures.

Dataset	Real	Fake	Forgery Method
CelebDFv1 [10]	409	795	FaceswapPro
CelebDFv2 [10]	590	5639	FaceswapPro
DeeperForensics1.0 [4]	50,000	10,000	DeepFake Variational Auto-Encoder (DF-VAE) [13]
FakeAVCeleb [3]	178	11,833	Faceswap [14], Faceswap GAN (FSGAN) [15], Wav2Lip [16]
DeepFakeMnist+ [2]	10,000	10,000	First Order Motion Model for Image Animation (FOMM) [17]
DeepfakeTIMIT [11]	320	640	faceswap-GAN [18]
FaceForensics++ [9]	1000	5000	Faceswap [19], Deepfakes [20], Face2Face [21], FaceShifter [22], NeuralTextures [23]
DeepFakeDetection [24]	363	3068	Faceswap
ForgeryNet [1]	99,630	121,617	ATVG-Net [25], BlendFace, DeepFakes, DeepFakes-StarGAN-Stack, DiscoFaceGAN [26], FaceShifter [22], FOMM [17], FS-GAN [15], MaskGAN [27], MMReplacement, SC-FEGAN [28], StarGAN-BlendFace-Stack, StarGAN2 [29], StyleGAN2 [30], Talking Head Video [31]
Patch-wise Face Image Forensics [12]	* 25,000	* 25,000	PROGAN [32], StyleGAN2 [30]

3. Troubles with Current Deepfake Traceability

4. The KCE-System

We assume that incorporating datasets that use the same forgery methods will beneficially enhance the model’s performance. Conversely, merging different datasets or dividing the similar dataset into separate subsets may adversely affect the model’s performance. Based on the above assumptions, we developed the K-means and Calinski Harabasz Evaluation System. For the sake of simplicity, we refer to it as the KCE-System for short.

The KCE-System incorporates unsupervised learning. The system divided the deepfake datasets into training sets and evaluation sets. Then it trains a deepfake recognition model using training sets and extracting high-dimensional vectors from the middle layer of the model. After dimensionality reduction, the system used the K-means clustering method to merge various deepfake datasets. Using these datasets, the system then trains the new Xception, F3-net, and ResNet models. The trained models are then used to extract 2048-dimensional or 512-dimensional values from the evaluation set as feature values. Finally, the system uses the Calinski Harabasz Index method on the feature values after dimensionality reduction to evaluate The model’s performance, as shown in Figure 2. Next, we will introduce several main parts of the system in detail.

Figure 2. Overview of the KCE-System. The proposed architecture consists of two parts: the cluster section and the evaluation section.

4.1. Feature Extractor

Theoretically, when a model reaches a high classification accuracy for various categories of deep fake data, the model can extract the corresponding deepfake feature. The Xception is a traditional CNN model based on separable convolutions with residual connections. The model has shown high accuracy when detecting deepfake videos. The training accuracy rate reaches 94%. We use it as the main Feature Extractor. We take out its 2048-dimensional data as the sample’s feature from the global pooling layer of Xception. The ResNet is an improvement over the traditional deep neural network architecture that solves the problem of vanishing gradients and allows the training of much deeper networks. Another notable model in facial forgery detection is the F3-Net. This model leverages frequency domain analysis and comprises two branches; one learns forgery patterns via Frequency-aware Image Decomposition, and the other extracts high-level semantics from Local Frequency Statistics. Given the widespread applicability of the ResNet model in various computer vision fields and the unique position of the F3-Net in the domain of deepfake detection, we also select these two models as Feature Extractors and test them on half of the test group. To avoid the interference of the model itself on the experimental results to the greatest extent.

4.2. Dimensionality Reduction and Clustering

In this field, clustering algorithms, such as K-means ^[39], Gaussian Mixture, and DBSCAN ^[40] are commonly used. However, the DBSCAN algorithm is ineffective in controlling the number of clusters formed. In our system, we need to control the number of clusters formed for easy comparison with the data merged by name. The Gaussian Mixture algorithm is mainly designed for non-spherical clusters, while we focus more on the distance between categories in feature space, which emphasizes spherical clustering. Therefore, we chose to use the K-means clustering algorithm in our system.

The K-means algorithm uses Euclidean distance for clustering, but it can fail in high dimensions, so a dimension reduction method must be used. PCA ^[41] and t-SNE ^[42] are two methods we utilized for comparison. PCA is stable but retains less information when reduced to two or three dimensions. When reducing dimensions to 64 using PCA, the interpretable variance contribution rate can be preserved at 95.2%. From Figure 3, it effectively preserves most of the information needed for clustering. The t-SNE supports low-dimensional reduction for visual analysis but has poor stability.

Figure 3. Illustration of dimensionality reduction using PCA. After using PCA to reduce the dimension, use the t-SNE method to reduce the dimension to two dimensions for display (Different colors indicate different forgery methods).

4.3 Selection of Evaluation Algorithms

Evaluating the performance of models trained with unreliably labeled or unlabeled data is difficult. We can not use precision and recall because we do not have a way to figure out whether each sample is classified correctly. To address this issue, we utilize the Calinski Harabasz Index ^[6], introduced by Calinski and Harabasz in 1974, as an effective evaluation method. This index is defined in Equation (1) as the ratio of the sum of between cluster dispersion and inter-cluster dispersion for all clusters. Therefore, the Calinski Harabasz Index can be used to evaluate the models, with higher scores indicating that the model performs better on the test datasets.

For a set of data of size , which has been clustered into clusters, the Calinski Harabasz score s is defined as the ratio of the between-cluster dispersion means and the within-cluster dispersion, as shown in Equation (1).

                                    (1)

where is trace of the between group dispersion matrix and is the trace of the within-cluster dispersion matrix defined by:

                       (2)

                        (3)

Here, represents the set of points in cluster, represents the center of cluster , represents the center of , and represents the number of points in cluster .

When using the Calinski Harabasz Index to evaluate clustering quality, it can be observed that the elbow points of the Calinski Harabasz Index tend to be around 3 or 4 of cluster number, as depicted in Figure 4. The results obtained from the Calinski Harabasz Index are consistent with the number of forged method categories in the actual evaluation set. This suggests that the Calinski Harabasz Index is a valuable method to assess the model’s ability to identify new categories of deepfakes.

Figure 4. Using Calinski Harabasz Index to evaluate its clustering quality, it can be found that its elbow point is about 3 to 4.

5. Experiment

In this section, we first introduce the overall experimental setup. Our equipment includes four NVIDIA GeForce2080Ti GPUs. We use PyTorch to train and evaluate models, OpenCV to image data preprocessing, and Scikit-learn algorithm library for data analysis. We extract 620,000 fake face images from 10 deepfake datasets and train 40 models, including 32 Xception, 4 F3-net, and 4 ResNet models. The entire data preparation and experimental process spanned approximately three months.

5.1. Data Dividing and Preprocessing

The researchers select 31 datasets labeled with forgery method names from CelebDF, DeeperForensics1.0, DeepFakeMnist+, FaceForensics++, ForgeryNet, and FakeAVCeleb; see Table 1 for details. The researchers use a random method to divide 31 deepfake categories into two sets, where the training set contains 27 categories, and the evaluation set contains four categories. The researchers repeat the above division four times to obtain four sets of training sets and evaluation sets. See Table 2 for details. The researchers extract the frame data of each category according to the instructions of the relevant dataset and use the face detection model Retinaface ^[43] to intercept the face area. Then, the researchers increase the side length of the area by a factor of 1.25. Finally, the researchers randomly select 20,000 fake faces of each category and save these images as test data in png format.

Table 2. The table displays four sets of experimental data, each containing four evaluation datasets, with the remaining 27 datasets designated for training purposes.

5.2. Merge Training Data Based on the Category Name

To verify our conjecture that there is large randomness in the naming of the forged methods in the deepfake dataset, we specially merged the training set data according to the principle of the same or close to the forged method names and used them as a control group. We use the merging rules see Table 3. The number of training set categories of the merged four groups are that Group 1, 3, and 4 have 19 categories, and Group 2 has 17.

Table 3. The researchers randomly sample corresponding proportions of data from the merged dataset and reassemble them into 20,000 images per category.

5.3. Merge Training Data Based on the Results of K-Means Clustering

One of the purposes of our experiment is to determine the appropriate dimensionality for K-means clustering to address this type of problem. We need to ensure that we do not lose too many classification features due to excessive dimensionality reduction, nor do we cause the K-means algorithm to fail due to excessive dimensionality. We use the PCA algorithm to reduce the Xception model's 2048-dimensional output to 128, 64, and 32 dimensions. We also reduce it to two dimensions using the t-SNE algorithm. For the F3-net and ResNet models, we only use the PCA algorithm to reduce the output feature value to 64 dimensions since we only need to verify that our method applies to these models.

In the previous section, we created training data for the control group based on name mergers. To facilitate comparison, we ensure that the number of categories of the experimental data for each group is identical. Therefore, we use the K-means clustering algorithm to cluster these training sets based on the specified number of clusters. Groups 1, 3, and 4 have 19 clusters, while Group 2 has 17 clusters.

5.4. Experimental Results

The researchers train Xception, F3-net, and ResNet models using training data merged by K-means clustering results and category names, respectively. For comparison, the researchers also train the same models using the original training set without merging. To obtain feature vectors for the validation set, we used these models as feature extractors and applied PCA to reduce them to 64 dimensions. The researchers then calculated the Calinski Harabasz Index. Please refer to Table 4 for the result.

Table 4. The Calinski Harabasz Index results. Italicized and underlined marks indicate the best result for that group of tests.

The Calinski Harabasz Index of the model trained on the data merged by K-means is 42.27% higher than that pooled by name. Furthermore, these scores are slightly higher than those directly using the original training set, even though the original set contains more data. At the same time, the Calinski Harabasz Index is also higher at 22.66% and 14.27% in F3-net and ResNet models. These prove an appropriate combination of deepfake datasets with similar features improves the model’s generalization in the unknown forgery categories.

Compared with the other three groups, the results of Group 2 are different. Furthermore, its Calinski Harabasz Index is lower than the training results on the original data. Because Group 2 has only 17 categories after the merger, with fewer training samples than other groups. More information loss can destroy the performance of the model.

6. Conclusions

The researchers prove the labels of various deepfake datasets contain many randomnesses. If researchers use more than two deepfake datasets, combining these datasets only based on forgery labels will hurt the model's performance. We propose K-means and Calinski Harabasz evaluation systems to evaluate the similarity of various deepfake datasets, laying the foundation for future researchers to use them comprehensively. The generalization ability of the deepfake recognition model in the face of new samples can be improved by merging datasets with high forgery feature similarity.

Our research revealed the arbitrariness of label naming in deepfake datasets and the resulting troubles in the traceability of forgery methods. There is still a long way to go to solve this problem completely. In addition, different image compression algorithms and image resolutions significantly impact the fake features of deepfake datasets, which will seriously interfere with the model’s extraction of fake features from deepfake datasets. We are committed to conducting further research to address these challenges effectively.

Furthermore, to ensure the healthy development of the field, we appeal to researchers and companies to standardize the label nomenclature of deepfake datasets.