The Same Name Is Not Always the Same: Correlating and Tracing Forgery Methods across Various Deepfake Datasets

Abstract

Deepfakes are becoming increasingly ubiquitous, particularly in facial manipulation. Numerous researchers and companies have released multiple datasets of face deepfakes labeled to indicate different methods of forgery. However, naming these labels is often arbitrary and inconsistent, leading to the fact that most researchers now choose to use only one of the datasets for research work. However, researchers must use these datasets in practical applications and conduct traceability research. In this study, we employ some models to extract forgery features from various deepfake datasets and utilize the K-means clustering method to identify datasets with similar feature values. We analyze the feature values using the Calinski Harabasz Index method. Our findings reveal that datasets with the same or similar labels in different deepfake datasets exhibit different forgery features. We proposed the KCE system to solve this problem, which combines multiple deepfake datasets according to feature similarity. We analyzed four groups of test datasets and found that the model trained based on KCE combined data faced unknown data types, and Calinski Harabasz scored 42.3% higher than combined by forged names. Furthermore, it is 2.5% higher than the model using all data, although the latter has more training data. It shows that this method improves the generalization ability of the model. This paper introduces a fresh perspective for effectively evaluating and utilizing diverse deepfake datasets and conducting deepfake traceability research.

1. Introduction

Facial recognition has become increasingly prevalent in recent years, with many applications utilizing it as the primary method for identity recognition. However, with the rapid development of deep learning-driven facial forgery technologies in recent years, such as deepfakes [1], there has been a rise in fraudulent practices within media and financial fields, which has sparked widespread social concern [2,3,4]. Consequently, there is a crucial need for the traceability of forged data.

Deepfake tracking methods can be broadly classified into traditional [5,6,7] and deep learning-based methods [8,9]. Traditional methods rely on techniques, such as image forensics and metadata analysis to detect signs of manipulation in a deepfake. These methods are based on analyzing the visual properties of an image or video, and they can include analyzing the distribution of colors, identifying inconsistencies in lighting and shadows, or detecting distortions in the image caused by manipulation. These traditional methods require extensive domain knowledge and specialized software to execute. On the other hand, deep learning-based methods rely on machine learning algorithms’ power to detect deepfakes. These methods train deep neural networks on large datasets of real and fake images or videos, and they can detect deepfakes by analyzing the patterns in the data. Deep learning-based methods are highly effective at detecting deepfakes, but they require large amounts of training data and computing resources to execute. This paper mainly conducts related research based on the latter method.

Tracing the source of deep forgery relies on identifying the forgery algorithms used. However, the category labels in deepfake datasets fundamentally differ from those in the general computer vision field. In typical computer vision datasets, such as the CIFAR [10], ImageNet [11], and MNIST [12], the category labels are objective and have real-world meaning. For instance, the labels for salamander and setosa are assigned by biologists based on the biological characteristics of these species, or humans can accurately recognize facial expressions such as anger or happiness, as shown in Figure 1. These labels remain unchanged despite variations in camera equipment, lighting conditions, and post-processing of images. However, humans cannot classify deepfake pictures visually, and the images can only be named based on their forgery method. The names given to the forgery methods by different producers are highly subjective and arbitrary, as shown in Table 1. Many “wild datasets” do not provide forgery method labels. Furthermore, subsequent operations such as image compression and format conversion [13] may significantly alter the forgery characteristics of the images.

Improving facial forgery recognition and tracking technology relies on collecting and utilizing as many facial forgery datasets as possible. These datasets include ForgeryNet [14], DeepfakeTIMIT [15], FakeAVCeleb [16], DeeperForensics-1.0 [17], and others. Additionally, numerous “wild datasets” are gathered from the Internet. However, these datasets are published by different institutions, use varying forgery methods, and have different naming conventions. In some cases, the exact generation algorithm is not provided. This situation leads some researchers to use only one dataset in their experiments. Dealing with those with similar or identical names can create challenges for users when multiple datasets are employed.

Measuring the relevance of each deepfake dataset is crucial. To address this problem, we use the Xception model [18] as a forgery feature extractor, which is commonly used in the deepfake recognition field. We train both multi-classification and binary classification models that map various deepfake images into the feature space illustrated in Figure 2.

After mapping the deepfake datasets to feature space, we use PCA for dimensionality reduction and the K-means method for clustering. We use these cluster datasets to retrain the Xception model. We also combine these deepfake datasets based on forgery method labels and use them to train another Xception model as a control group. We perform a series of experiments on the test data using these models and use the Calinski Harabasz Index [19] as a measure to judge the performance of the models. To improve the credibility of the experimental results, we also repeat some experiments on The Frequency in Face Forgery Network (F3-Net) [9] and Residual Neural network (ResNet) [20].

Our main contributions are summarized as follows:

• We point out that the forgery category labels in the deepfake dataset lack objectivity. Our experiments prove that some forgery category labels of the same name differ significantly across different datasets.
• We establish the KCE-System. It is a deepfake dataset similarity evaluation index system that provides a measure of the similarity between different datasets and lays the foundation for subsequent researchers to use these datasets comprehensively.
• Our experiments confirm that when the forgery method of the deepfake dataset is unknown, the model can achieve better generalization performance by training on datasets that are merged based on closer feature distances.

2. Related Works

2.1. Deepfake Datasets

Numerous deepfake datasets have been created by researchers and institutions, including FaceForensics++ [21], Celeb-DF [22], DeepFakeMnist+ [15], DeepfakeTIMIT [1], FakeAVCeleb [16], DeeperForensics-1.0 [17], ForgeryNet [14], and Patch-wise Face Image Forensics [23]. These datasets cover various forgery methods, have significant data scales, and are widely used. Please refer to Table 1 for more details.

Table 1. Common deepfake datasets, the symbol * represents the number of pictures.

Dataset	Real	Fake	Forgery Method
CelebDFv1 [22]	409	795	FaceswapPro
CelebDFv2 [22]	590	5639	FaceswapPro
DeeperForensics1.0 [17]	50,000	10,000	DeepFake Variational Auto-Encoder (DF-VAE) [24]
FakeAVCeleb [16]	178	11,833	Faceswap [25], Faceswap GAN (FSGAN) [26], Wav2Lip [27]
DeepFakeMnist+ [15]	10,000	10,000	First Order Motion Model for Image Animation (FOMM) [28]
DeepfakeTIMIT [1]	320	640	faceswap-GAN [29]
FaceForensics++ [21]	1000	5000	Faceswap [30], Deepfakes [31], Face2Face [32], FaceShifter [33], NeuralTextures [34]
DeepFakeDetection [35]	363	3068	Faceswap
ForgeryNet [14]	99,630	121,617	ATVG-Net [36], BlendFace, DeepFakes, DeepFakes-StarGAN-Stack, DiscoFaceGAN [37], FaceShifter [33], FOMM [28], FS-GAN [26], MaskGAN [38], MMReplacement, SC-FEGAN [39], StarGAN-BlendFace-Stack, StarGAN2 [40], StyleGAN2 [41], Talking Head Video [42]
Patch-wise Face Image Forensics [23]	* 25,000	* 25,000	PROGAN [43], StyleGAN2 [41]

Table 1. Common deepfake datasets, the symbol * represents the number of pictures.

Dataset	Real	Fake	Forgery Method
CelebDFv1 [22]	409	795	FaceswapPro
CelebDFv2 [22]	590	5639	FaceswapPro
DeeperForensics1.0 [17]	50,000	10,000	DeepFake Variational Auto-Encoder (DF-VAE) [24]
FakeAVCeleb [16]	178	11,833	Faceswap [25], Faceswap GAN (FSGAN) [26], Wav2Lip [27]
DeepFakeMnist+ [15]	10,000	10,000	First Order Motion Model for Image Animation (FOMM) [28]
DeepfakeTIMIT [1]	320	640	faceswap-GAN [29]
FaceForensics++ [21]	1000	5000	Faceswap [30], Deepfakes [31], Face2Face [32], FaceShifter [33], NeuralTextures [34]
DeepFakeDetection [35]	363	3068	Faceswap
ForgeryNet [14]	99,630	121,617	ATVG-Net [36], BlendFace, DeepFakes, DeepFakes-StarGAN-Stack, DiscoFaceGAN [37], FaceShifter [33], FOMM [28], FS-GAN [26], MaskGAN [38], MMReplacement, SC-FEGAN [39], StarGAN-BlendFace-Stack, StarGAN2 [40], StyleGAN2 [41], Talking Head Video [42]
Patch-wise Face Image Forensics [23]	* 25,000	* 25,000	PROGAN [43], StyleGAN2 [41]

2.2. Deepfake Identification and Traceability

2.2.1. Methods Based on Spectral Features

Many scholars consider upsampling to be a necessary step in generating most face forgeries. Cumulative upsampling can cause apparent changes in the frequency domain, and minor forgery defects and compression errors can be well described in this domain. Using this information can identify fake videos. Spectrum-based methods have certain advantages in generalization because they provide another perspective. Most existing image and video compression methods are also related to the frequency domain, making the method based on this domain particularly robust.

Chen et al. [44] proposed a forgery detection algorithm that combines spatial and frequency domain features using an attention mechanism. The method uses a convolutional neural network and an attention mechanism to extract spatial domain features. After the Fourier transform, the frequency domain features are extracted, and, finally, these features are fused for classification. Qian et al. [9] proposed a network structure called F3-Net (Frequency in Face Forgery Network) and designed a two-stream collaborative learning framework to learn the frequency domain adaptive image decomposition branch and image detail frequency statistics branch. The method has a significant lead over other methods on low-quality video. Liu et al. [45] proposed a method based on Spatial Phase Shallow Learning (SPSL). The method combines spatial images and phase spectra to capture upsampled features of facial forgery. For forgery detection tasks, local texture information is more critical than high-level semantic information. By making the network shallower, the network is more focused on local regions. Li et al. [46] proposed a learning framework based on frequency-aware discriminative features and designed a single-center loss function (SCL), which only compresses the intra-class variation of real faces while enhancing the inter-class variation in the embedding space. In this way, the network can learn more discriminative features with less optimization difficulty.

2.2.2. Methods Based on Generative Adversarial Network Inherent Traces

Scholars suggest that fake faces generated by generative adversarial networks have distinct traces and texture information compared to real-world photographs.

Guarnera et al. [47] proposed a detection method based on forgery traces, which uses an Expectation Maximization algorithm to extract local features that model the convolutional generation process. Liu et al. [48] developed GramNet, an architecture that uses global image texture representation for robust forgery detection, particularly against image disturbances such as downsampling, JPEG compression, blur, and noise. Yang et al. [49] argue that existing GAN-based forgery detection methods are limited in their ability to generalize to new training models with different random seeds, datasets, and loss functions. They propose DNA-Det, which observes that GAN architecture leaves globally consistent fingerprints, and model weights leave varying traces in different regions.

2.3. Troubles with Current Deepfake Traceability

Methods based on frequency domain and model fingerprints provide traceability for different forgery methods. Although researchers claim high accuracy rates in identifying and tracing related forgery methods, they typically only use a specific dataset for research. This approach reduces the comprehensiveness of traceability and the model’s generalization ability. Therefore, researchers need to consider the similarity and correlation between samples in each dataset to make full use of these datasets.

However, this presents a significant challenge. Unlike typical computer vision datasets, deepfake datasets’ labels are based on technical methods and forgery patterns rather than human concepts, making it impossible for humans to identify and evaluate them. The more severe problem is that the labels of forgery methods used in various deepfake datasets are entirely arbitrary. Some labels are based on implementation technology, while others are based on forgery modes. For example, many datasets have the label “DeepFakes”. The irregularity and ambiguity of these labeling methods make it difficult to utilize the forged data of various deepfake datasets fully. Additionally, some deepfake datasets do not indicate specific forgery methods, such as “wild datasets”.

3. Research Methods

3.1. K-Means and Calinski Harabasz Evaluation System

We trained an Xception model as a feature extractor using various deepfake datasets and real datasets as training sets. When examining different deepfake datasets in feature space, we observe that specific forgery methods are clustered together. In contrast, some forgery methods with similar names are separated, as shown in Figure 2. For example, one of the FOMM forgery methods is very close to the FaceSwap method but far from the other FOMM forgery methods. It shows that the forgery methods with the same name have a significant feature gap in different datasets, and different forgery methods will have relatively similar features. The same trend can be seen in the Cosine Similarity results in Figure 3. In order to evaluate the similarity between different forgery methods across various datasets. 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. We developed the K-means and Calinski Harabasz Evaluation System based on the above assumptions. 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. The system then trains the new Xception, F3-net, and ResNet models using these datasets. 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 4. Next, we will introduce several main parts of the system in detail.

3.2. 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. We use the trained deepfake recognition model as a feature extractor, as the accuracy of these models in deepfake multi-classification tasks can reach more than 90%. For a comprehensive evaluation, we provide several representative models with different sizes.

The Xception [18] is a traditional CNN model based on separable convolutions with residual connections. The model has shown high accuracy when detecting deepfake videos. In terms of the training process of the feature extractor, the forgery method indicated in each dataset is used as a pseudo-labelling for multi-class training on the Xception. The training accuracy rate reaches 94%, and the model converges after three rounds of training. We use the trained model extract feature on the data of 27 categories of deepfake datasets. We take out its 2048-dimensional data as the sample’s feature from the global pooling layer of Xception. Considering the trade-off between performance and efficiency, we select Xception as the baseline model.

The ResNet [20] 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. One of the main advantages of ResNet is its ability to handle deeper architectures, which leads to better accuracy in image classification tasks. Another notable model in facial forgery detection is the F3-Net, as proposed in [9]. This model leverages frequency domain analysis and comprises two branches, one focused on learning subtle forgery patterns via Frequency-aware Image Decomposition (FAD) and the other aimed at extracting high-level semantics from Local Frequency Statistics (LFS). Extensive experiments have demonstrated the effectiveness of the F3-Net in identifying low-quality forgery videos. 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 evaluation models 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.

3.3. Dimensionality Reduction and Clustering

In this field, clustering algorithms, such as K-means [50], Gaussian Mixture, and DBSCAN [51] 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. Two methods we utilized for comparison are PCA [52] and t-SNE [53]. 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 5, we can see that it effectively preserves most of the information needed for clustering. The t-SNE supports low-dimensional reduction for visual analysis, but it has poor stability.

We utilize five different dimensionality reduction parameters to determine the most appropriate clustering dimension. We apply the t-SNE algorithm to reduce the high-dimensional feature data to 2 dimensions and use the PCA algorithm to reduce the dimensionality to 32, 64, and 128 dimensions. We also keep the 2048 dimensional original features without applying any dimensionality reduction algorithm. We then performed K-means clustering on each of these dimensions individually.

3.4. Selection of Evaluation Algorithms

We select four categories of deepfake datasets not involved in the training and clustering process as evaluation sets. We extract Xception, ResNet, and F3-net models’ global pooling layer output and use the PCA algorithm reduces the data to 128 dimensions. An example of the results in Figure 6, demonstrating a clear distinction between the four unknown deepfake categories. This figure indicates that our model has indeed learned the relevant characteristics for identifying deepfakes.

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 [19], 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 E of size $n_E$, which has been clustered into k 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).

$$s=\frac{tr\left(B_k\right)}{tr\left(W_k\right)}\times \frac{n_E-k}{k-1}$$

(1)

where $tr\left(B_k\right)$ is trace of the between group dispersion matrix and $tr\left(W_k\right)$ is the trace of the within-cluster dispersion matrix defined by:

$$B_k=\sum _{q=1}^k{n}_q(c_q-c_E){(c_q-c_E)}^T$$

(2)

$$W_k=\sum _{q=1}^k\sum _{x\in C_q}^{}(x-c_q){(x-c_q)}^T$$

(3)

Here, $C_q$ represents the set of points in cluster q, $c_q$ represents the center of cluster q, $c_E$ represents the center of E, and $n_q$ represents the number of points in cluster q.

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 7. 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. When other training parameters remain the same, if a model’s performance is outstanding, it indicates that the quality of the training set is excellent, with fewer incorrect labels. In other words, we effectively improve the reliability of these classification labels in the training set. Therefore, the Calinski Harabasz Index can effectively evaluate the correlation of these unreliable classification labels in our system.

4. 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 3 months.

4.1. Data Dividing and Preprocessing

We select 31 datasets labeled with forgery method names from CelebDF, DeeperForensics-1.0, DeepFakeMnist+, FaceForensics++, ForgeryNet, and FakeAVCeleb, see Table 1 for details. We 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. We repeat the above division four times to obtain four sets of training sets and evaluation sets. See

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

Group	Evaluation Datasets
1	Faceforensics++_FaceShifter, FakeAVCeleb_FaceSwap, ForgeryNet_ATVG-Net, ForgeryNet_FOMM
2	DeeperForensics_DF-VAE, Faceforensics++_Face2Face, FakeAVCeleb_Wav2Lip, ForgeryNet_DiscoFaceGAN
3	DeepfakeTIMIT_FaceSwap-GAN, ForgeryNet_BlendFace, ForgeryNet_StarGAN2, Patch-wise-Face-Image-Forensics _PROGAN
4	CelebDFv2_FaceSwapPRO, Faceforensics++_ NeuralTextures, ForgeryNet_FS-GAN, ForgeryNet_Talking Head Video

for details.

We extract the frame data of each category according to the instructions of the relevant dataset and use the face detection model Retinaface [54] to intercept the face area. Then, we increase the side length of the image by a factor of 1.25. Finally, we randomly select 20,000 fake faces of each category and save these images as test data in png. format.

4.2. Merge Training Data Based on the Category Name

In order 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 following merging rules.

(a)

Merge the FaceSwapPRO category in the CelebDFv1 dataset and the FaceSwapPRO category in the CelebDFv2 dataset.

(b)

Merge the FOMM category in the DeepFakeMnist+ dataset and the FOMM category in the ForgeryNet dataset.

(c)

Merge the FaceSwap-GAN category in the DeepfakeTIMIT dataset, the DeepFakeDetection FaceSwap category and the FaceSwap category in the FaceForensics++ dataset, and the faceswap category in the FakeAVCeleb dataset.

(d)

Merge the DeepFakes category in the Faceforensics++ dataset and the DeepFakes category in the ForgeryNet dataset.

(e)

Merge the FSGAN category in the FakeAVCeleb dataset and the FS-GAN category in the ForgeryNet dataset.

(f)

Merge DeepFakes-StarGAN-Stack category, StarGAN-BlendFace-Stack category and StarGAN2 category in ForgeryNet dataset.

(g)

Merge the StyleGAN2 category in the ForgeryNet dataset and the STYLEGAN2 category in the Patch-wise Face Image Forensics dataset.

We randomly sample corresponding proportions of data from the merged dataset and reassemble them into 20,000 images per category. The number of training set categories of the merged four groups is that Group 1 has a total of 19 categories, Group 2 has a total of 17 categories, Group 3 has a total of 19 categories, and Group 4 has a total of 19 categories.

4.3. 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. Since we chose the Xception model as the baseline, we use the PCA algorithm to reduce the 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. Finally, we use the results of the K-means clustering algorithm to combine the training set.

4.4. Experimental Results

We train Xception, F3-net, and ResNet models using training data merged by K-means clustering results and category names, respectively. For comparison, we 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. We then calculated the Calinski Harabasz Index. Please refer to

Table 3. The Calinski Harabasz Index results. The bold type indicates the best result for that group of tests.

Model	Train Data Merge by	Group 1 CH	Group 2 CH	Group 3 CH	Group 4 CH	Avg CH
Xception	Without merging	128.02825	117.448499	68.6994684	93.5723306	101.937137
Xception	Name	84.0837009	73.8172086	74.579957	61.2651927	73.4365148
Xception	K-means on 2048D	124.241305	105.070655	76.2218761	84.212058	97.4364735
Xception	K-means on t-SNE 2D	103.627829	87.1461055	66.6143003	76.5264273	83.4786656
Xception	K-means on PCA 64D	137.241584	101.192327	85.2535376	94.2137508	104.4753
Xception	K-means on PCA 128D	101.197038	101.502163	74.8441997	86.6358341	91.0448087
Xception	K-means on PCA 32D	114.247635	89.1934801	62.3932779	75.9596147	85.4485019
F3-net	Name			62.6592813	65.6510862	64.1551837
F3-net	K-means on PCA 64D			85.361067	72.018708	78.6898875
ResNet	Name			42.895651	47.9716533	45.4336522
ResNet	K-means on PCA 64D			49.7529116	54.0786263	51.915769

for the result.

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.

The Calinski Harabasz Index of merging by names is lower than by various cluster-based and original training sets, indicating significant differences in the characteristics of these same-name forgery methods. Merged by name harms the model.

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.

5. Conclusions

This article starts with the traceability requirements of the deep forgery method. When using multiple deepfake datasets, we found many different deepfake datasets using the same or similar label names. Confusion arises in how to use these datasets comprehensively.

We leverage the Xception model to extract fake features from the deepfake dataset. Subsequently, PCA and t-SNE methods are employed to reduce dimensionality and perform K-means clustering. Then, combine the datasets based on the clustering results, and use the combined data to train Xception, F3-net, and ResNet models, respectively. Finally, we use these models to extract features from the evaluation set and evaluate the generalization of these models using the Calinski Harabasz index as an evaluation metric. Our contributions are mainly three-fold:

• We 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 performance of the model.
• 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.
• We prove that 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 is only a helpful exploration for entirely using various deep forgery datasets from the source of deep forgery methods. We mainly 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, and pose a significant challenge to the identifiability and traceability of deepfake datasets. We are committed to conducting further research to address these challenges effectively.

To ensure the healthy development of the field, research institutions and universities should standardize the label nomenclature of deepfake datasets. Additionally, legislation should require digital watermarking and blockchain technology to trace deepfake content to its source accurately. Our research is a helpful exploration of the use of various deep forgery datasets, and we hope it will inspire future work in this field.