Deep Learning has greatly impacted society, leading to impressive advancements in various fields. However, its use can also have negative consequences, for example, the creation of deepfakes. Deepfakes are manipulated images or videos that depict subjects in ways they never actually were, which can harm reputations or manipulate reality. Indeed, although deepfakes have numerous potential applications in the fields of entertainment, art, and education, they also pose significant security and ethical risks. For this reason, it is crucial to continue the development of robust deepfake detection methods to counteract such a threat. To tackle this problem, researchers have developed deepfake detection techniques, which are usually based on deep learning as well. These methods try to identify any traces introduced during the manipulation process, but they require large amounts of data for training. Furthermore, deepfakes are generated by resorting to different typologies of techniques and/or procedures (often even unknown) that emerge almost daily, so it is not possible to follow each methodology and consequently to re-adapt the training phase. On this basis, to have more effective deepfake detectors, researchers aim for a system that can generalize the concept of deepfakes and identify them regardless of the manipulation technique used, even if it is novel and not present in the training data. During training, a huge amount of heterogeneous data are needed to provide to the models in order for them to see enough forms of deepfakes to stimulate them to abstract and generalize. A comparison was made among different deep learning architectures in order to validate their generalization capabilities, specifically against deepfake videos. It is worth saying that techniques used to manipulate videos do not necessarily introduce the same anomalies and features that can be embedded when tampering with still images.

2. Experiments

2.1. General Setup

2.2. Single Method Training

Video Manipulation Methods	Training Frames	Test Frames
0 (Pristine)	118,255	47,369
1 (FaceShifter)	13,337	1889
2 (FS-GAN)	48,122	8732
3 (DeepFakes)	8550	1157
4 (BlendFace)	9827	1335
5 (MMReplacement)	270	115
6 (DeepFakes-StarGAN-Stack)	3610	509
7 (Talking-Head Video)	26,338	2199
8 (ATVG-Net)	37,449	5383

2.3. Multiple Methods Training

Video Manipulation Categories	Training Frames	Test Frames
0 (Pristine)	118,255	47,369
ID-Replaced	83,716	13,737
ID-Remained	63,787	7582

Figure 3 shows the accuracies achieved by the three considered models trained in the Single Method Training setup, presented in the previous section, with respect to each of the methods comprised within the test set. Looking at the accuracies of the three models, it can be pointed out that the EfficientNetV2-M and the Swin-Small maintain results often above 80%

in correspondence of test frames manipulated with the same methods used in the training set (as expected) and, at the same time, obtain a certain degree of generalization. In fact, the same models sometimes succeed in detecting frames manipulated with methods unseen during training, although only reaching values of accuracy that are quite limited. The case of method number 5 (MMReplacement) is rather anomalous, though the detection percentage is often very high indeed; this behaviour is probably induced by the low number of available examples (see Table 1).

On the contrary, it can be easily noticed that, in all the cases, the ViT-Base is substantially unable to learn in the presence of relatively few training images. In fact, for instance, by training the model on methods 3, 4, 5, and 6 and then testing it on the test set, it is evident that the model is substantially underfitting and practically unusable compared to the two others taken into consideration. Interestingly, the Swin Transformer, although also based on the attention mechanism, is not particularly affected by this phenomenon and instead succeeds in obtaining competitive results in all contexts. This probably lies in the hierarchical nature that emulates the convolutional layers of traditional CNNs and thus allows it to exploit implicit inductive biases better. Good performances are preserved, in any case, with respect to pristine frame detection. In this setup, the architecture based on a convolutional network seems to prove more capable of generalization. The accuracies obtained from the three models are also shown in the confusion matrices in Figure 4 where all previously commented trends are reconfirmed again.

The behaviour of the three networks is now analysed in the second considered setup, namely Multiple Methods Training, and corresponding results are shown in Figure 5. In this case, the datasets are composed of frames extracted from videos manipulated by not only one method, so the models will have more difficulty focussing on specific artifacts and be forced to generalize. In this setup, the situation is significantly different from the previous one. Surprisingly, the classic Vision Transformer, which previously struggled to train effectively, is now the only model capable of generalizing well to frames that have been manipulated using techniques that were not present in the training data. This result probably stems from the fact that the training set consists of significantly more images than in the previous setup and it is strongly in line with what is presented in ^[1]. This particular architecture shows in many contexts a major need for data and resources which, when available, enable it to achieve very competitive results. In this case, the confusion matrices (see Figure 6) clearly show the greater generalization capacity of the Vision Transformer although at the expense of more false positives. In fact, the “pristine” class is less accurately classified by this latter. This may be a problem since in a real-world context researchers may want to reduce as much as possible the number of false alarms, in particular, if the system is fully automated.

To further evaluate the generalization capability of the trained models researchers also tested them in a cross-dataset context. In particular, researchers considered the three architectures trained on videos manipulated with ID-Remained or with ID-Replaced methods (ForgeryNet dataset) and tested on the well-known DFDC Preview test set. Researchers report the AUC values of the trained models compared with previous works in the literature. This is probably the most challenging scenario since both the contexts and the manipulation methods are significantly different, and indeed the performances of the models are pretty low. In particular, the EfficientNets are totally incapable of detecting these deepfakes with a very low AUC value. On the other hand, attention-based methods manifest better performances even if, as expected, they are worse than other, more complicated and articulated, methods in the literature. The Swin-Small trained on the videos manipulated with ID-Replaced methods perform pretty well with an AUC of 71.2%, demonstrating a good level of generalization. Again, also in this context, it seems that attention may be the key to achieve better generalization performances while the considered CNN is in any case too tied with the methods seen during training. Furthermore, the trained models which achieve better performances are the ones trained on a more complete and heterogeneous dataset, namely the videos manipulated with ID-Replaced methods, highlighting again the need for these architectures for a huge amount of data.