Securing Social Media Against Deepfakes using Identity, Behavioral, and Geometric Signatures

Traditional deepfake detection methods mainly relied on handcrafted features, specifically designed to exploit signature vulnerabilities introduced by the generative algorithms. These features targeted inconsistencies induced by the deepfake generation algorithms, for instance, the authors of the study [15] revealed that deep fake videos often lack natural physiological signals like regular eye blinking and consistent head movements, which could serve as indicators for deep fake detection. In another study [16], the authors identified visual anomalies, such as changes in eye color, unconvincing reflections, and discrepancies in eye and tooth details, and used them to identify deepfakes. Similarly, the method proposed in [17] exploited distinct facial expressions. In [18], the method capitalized on generator limitations, while [19] focused on subtle variations induced by the heartbeat on the face to discern video authenticity. Though these techniques were effective in their time, their dependence on specific cues made them vulnerable to rapid advancements in deepfake technology. In contrast, some other works concentrate on methods for creating interpretable deep models through the use of self-explanatory mechanisms [20]. Instead of relying solely on raw inputs, these approaches used ”units of explanation,” such as abstract notions, to convey interpretability. This strategy was used for basic concept acquisition by Alvarez et al. in [20], which was further used for case-based reasoning and prototype learning by Kim et al. [21]. Recently, the author of [22] proposed ProtoPNet, which uses prototype learning to create predictions based on resemblance to class-specific image patches. However, although significant progress has been made in the development of intricate video classification models for explainable solutions (such as 3D CNNs, TSNs, and TRNs [23, 24], there is a lack of interpretability in complex deepfake detection methods encompassing recent deepfake forgeries.

Deep learning approaches for deepfake detection offer more robust and adaptable solutions. Unlike traditional handcrafted features, neural networks are able to learn the complex patterns and subtle inconsistencies in real and fake faces, surpassing the limitations of manually engineered features. These solutions can be broadly categorized into frame-based, temporal, and spatial-temporal approaches. In frame-based deepfake detection,  [25] proposed a deepfake detector that identified blending imperfections for video distortion detection. The author of the study [26] presented a CNN-based model for detecting deformation artifacts. Similarly, the presented framework in [27] analyzed video segments to capture visual inconsistencies. In contrast, in temporal analysis, [28] combined CNNs and LSTMs to analyze blinking patterns and temporal irregularities. The author of [29] used a spatio-temporal network for deepfake manipulated artifact detection. On the other hand, some studies developed lightweight deepfake detector models that can be used in computational intensive devices. For instance, the author of the study [30] presents ShallowNet based deepfake detector which excelled at detecting GAN-generated images even at low resolution. In another study [31], Pellcier et.al., introduces a unified framework based on prototype learning for Deepfake Detection (PUDD), which uses similarity-based detection to identify deepfakes by comparing input data to known prototypes. The presented PUDD framework outperforms the existing methods by achieving 95.1% accuracy on the Celeb-DF dataset. However, the presented solution was not tested against generalization across diverse datasets. Additionally, the model’s effectiveness could be affected by sensitivity to prototype selection. Similarly, the study [32] investigates the potential of deep neural network fusion for effective deepfake detection. The author emphasize the structural ability of various neural networks to capture distinct features for deepfake detection. For this, the author presents a multi-branch and multi-level architecture that separates fixed and adaptive knowledge from pre-trained networks which enhance the detection on low-power devices such as mobile devices. However, the generalization ability of this solution remains untested. In the next subsection, we explained and discussed the research conducted towards generalization of the deepfake detection.

Generalization ability remains a critical hurdle in deepfake detection, affecting the performance of the deepfake detectors from adapting to unseen data. Despite this significant challenge, research specifically addressing generalization ability within this domain remains limited. Early attempts like FWA [33] exploited resolution differences between forged faces and backgrounds for detection. Recent studies, however, demonstrate significant progress towards enhanced generalizability. Face X-ray [34] focuses on identifying blending boundary artifacts, while SPSL [35] utilizes phase spectrum analysis in a frequency-based approach. Lip-Forensics [36] uses spatio-temporal networks to detect unnatural mouth movements, and SRM [37] analyzes high-frequency noise for broader detection capabilities. In recent study [38], The author Yan et.al., tackles the generalization issue in deepfake detection, where performance suffers from mismatched training and testing data distributions. The authors present the Latent Space Data Augmentation (LSDA) detector, which improves decision boundaries by simulating variations of forgery features in the latent space. The presented LSDA outperformed the existing detectors across multiple benchmarks. However, LSDA’s reliance on effective simulation techniques may restrict its applicability to real-world scenarios with novel unseen deepfake forgeries. In another study [39], the authors propose the Generalized Multi-Scenario Deepfake Detection framework (GM-DF), which trains models on multiple datasets and utilizes a hybrid expert modeling approach along with a domain-aware meta-learning strategy. Although the framework shows promise in enhancing generalization, it may experience rapid declines in accuracy due to dataset discrepancies. Although these studies enhance the generalization of deepfake detection methods, their reliance on predetermined forgery patterns and consideration of the entire feature space makes them susceptible to disruption from irrelevant factors like background and identity.

Therefore, collectively both handcrafted and deep learning approaches offer distinct advantages: dataset oriented deepfake detection for handcrafted features and generalization ability for deep neural network. However, their fusion within a single framework holds significant promise for achieving both in-dataset and generalized deepfake detection. This strategy has been explored in few studies, some of which are discussed in the next subsections, with the aim of enhancing the generalization ability for deepfake detection systems.

Given a video $V$ consisting of $N$ frames, reshaped into $K$ batches ${S}^{B\times C\times W\times H}_{K}$, where $B$, $C$, $W$ and $H$ represent batch size, frame channels, width and height, respectively. A Multi-task Cascaded Convolutional Network  [40] is then applied on each batch $S^{B\times C\times W\times H}_{k},\text{ for }k=1,2,3,\ldots,K-1$ to get the coordinates of faces in the input frames. Batches are reshaped into a sequence of frames ${V}^{N\times C\times W\times H}$ along with the coordinates of the detected faces. To ensure the quality of resulting cropped faces for the subsequent steps, we set $120\times 120$ as the acceptable facial dimensions required by the MTCNN, in the input video frames. Similarly, in the video with multiple faces, the face with larger dimensions is selected. After the face detection, faces are cropped and rescaled to $224\times 224$ to get ${V}^{N\times C\times W\times H}_{224\times 224}$, which is used for behavioral, geometric and deep identity features extraction.

We argue that a holistic analysis of a face in a given video is more robust and generalized approach than a single cue based analysis. Based on our hypothesis and reported experimental results, we proposed a multidisciplinary DBaG descriptor composed of behavioral, facial geometry, and identity signature for deepfake detection. The visualization of DBaG features descriptor is given in Figure 2 and the composition of the BDaG is visualized in Figure 3. The DBaG descriptor holistically analyze a given face for inconsistencies in facial behavior and face geometry along with relative identity information. Furthermore, the behavior and geometric features in DBaG descriptor are handcrafted, resulting in interpretability, while the identity features are extracted using deep models to detect low-level details like texture and skin color tone, and hence improve generalizability. A detailed discussion on each component of the DBaG descriptor is given in the subsequent sections.

To perform classification in an efficient manner, we designed a deep learning classifier DBagNet as shown in Figure 4, which employs residual learning alongside both spatial and temporal domains. This architecture incorporates attention-based mechanisms and pooling techniques, drawing on prior research [45, 46] to create a model that can effectively generalize across manipulated media.

To evaluate the model, the testing samples are passed through the trained network, generating embeddings for each slice. We implement a similarity comparison approach from the reference set for the labels prediction of test embeddings. We calculate the distance of each test embedding from the embeddings of reference set to create a distance matrix. The final label of test embedding is predicted based on the majority labels of m closest neighbors in the reference set.

Once the training process is complete, we generate embeddings for each sample in the training set, referred to as the reference set $E_{\text{ref}}$. $E_{\text{ref}}$ is stored as embeddings, labels pairs $\{(e_{1},y_{1}),(e_{2},y_{2})\ldots,(e_{n},y_{n})\}$, for n reference samples. These embeddings form a basis against which new (unseen) test samples are compared, providing a fixed point of reference for label prediction. For each sample in the test set, the model computes a new embedding $e_{test}$ by passing it through the trained DBagNet. This embedding represents the learned features of the test sample in the same embedding space as the reference set, capturing characteristics that differentiate real from fake data.To determine the label of $e_{test}$, we calculate its Euclidean distance from each reference embedding $e_{i}$ in $E_{ref}$, providing a measure of similarity. The Euclidean distance $d_{i}$ between $e_{test}$ and $e_{i}$ is given by:

where j indexes the dimensions of the embedding vectors. This step produces a set of distances ${d_{1},d_{2},d_{3},...,d_{n}}$ that indicate how close the test embedding is to each reference sample. To assign a label to $e_{test}$, we rank the distances ${d_{1},d_{2},d_{3},...,d_{n}}$ in ascending order and identify the $m$ smallest distances, representing the nearest neighbors of $e_{test}$ in the embedding space. The label is then determined by a majority vote among the labels of these nearest neighbors. Formally, the predicted label is given by:

where ${i_{1}},y_{i_{2}},\ldots,y_{i_{k}}$ are the indices of the $m$ nearest embeddings in $E_{ref}$ and $M_{vote}$ selects the label that appears most frequently among these neighbors. This majority voting process ensures that the final prediction takes into account multiple neighbors, providing robustness against outliers and minor variations in the embedding space.

We evaluate the feasibility and superiority of the proposed framework on six large-scale datasets of manipulated samples from recent GAN’s and Diffusion models, confirming a well-suited evaluation to assessing the model’s efficacy for generalization on new and old types of deepfakes. Short description of each dataset is given below.

In our evaluation, we used Accuracy (ACC), equal error rate (EER) and area under the curve (AUC) metrics. A random split of 80%:20% is used for training and testing, respectively for DFDC, DFD, WLDR, and NVFAIR as these datasets have enough videos for each identity. While for the FF++ and CelebDF datasets we followed a different approach for train/test split, similar to [52], because these datasets have a limited number of videos per identity. Therefore, we chose the training and testing subsets from each video as follows: the first 20% of the video segment is used for training, the last 20% for testing, and the remaining video is ignored to prevent overlapping.

To evaluate the performance of the proposed framework, we tested our framework on the two most common types of deepfake video i.e., face swapping and expressions swapping. A detailed discussion on the obtained results is presented in subsequent sections.

The proposed framework is compared with current state-of-the-art (SOTA) techniques to analyse its effectiveness over various datasets. We compared the proposed framework with PWL [17] and AnB [52] on WLDR dataset while 2-stream [53], XceptionNet [48],Head Pose [15], MesoNet [54], DSP-FWA[10], AnB [52], MAT [55], SLADD [56], Face-x-ray [57], PCL+12G [58], and SBI [59] on the rest of the datasets. Table III gives a comparative performance analysis. As seen, the proposed framework outperform PWL and AnB on the WLDR dataset, achieving an AUC of 99.29. It also achieves an AUC of 99.41 on FF++, 96.78 on DFDC, and 99.43 on the CelebDF dataset, outperforming other methods on DFDC and CDF datasets. The proposed method demonstrated strong performance, achieving AUC scores of 99.41% on the FF++ dataset and 92.12% on the DFD dataset, positioning it as the second-highest performer and closely matching the leading results of 99.64% and 93.2% attained by SBI and AnB, respectively.

To analyze the generalizability of the proposed framework on unseen type of deepfakes for the seen identities, we compared the performance against the SOTA on FaceShifter(FST), FaceSwap(FS) and DeepFakes(DF) manipulation classes of FF++. The model was trained on each of the three types of deepfakes (FST, FS, and DF) and tested on the test set of the same and the other two subsets. It can be observed in the Table IV and Figure 5 that the mean AUC achieved by our framework is significantly higher compared to the SOTA, with the mean AUC gains of 6.97, 12.39 and 16.08 respectively. The AUC gain for DF training is comparatively low due to the low performance on FST as shown in Figure 5 $(c)$. The reason for this performance drop is the difference in quality and the nature of these two manipulation algorithms. Videos generated with DF do not have natural eye blinking and lip movements, while the videos generated with FST have more realistic behavioral characteristics, making them more difficult to detect.

To analyze the generalizability of the proposed framework, we performed a comparative analysis with the SOTA by training the model on FF++ and testing on CelebDF, DFD and DFDC datasets. To conduct this experiment, the proposed framework was trained on the FF++(C23) and tested on test sets of other three datasets. We selected SOTA methods for comparison, including XceptionNet [48], MLDG [64], MAT [55], DCL [62], EN-b4 [60], GFF [61] and IID [63]. It is shown in Table V, the proposed model achieves superior performance over the SOTA.

To analyze the effectiveness of extracted features, we performed an ablation study to evaluated our framework with the different combinations of feature sets on the DFDC, WLDR, FFF++ face swap, FF++ expression swap and NVFIAR datasets. The quantitative results on each dataset are presented in Table VI. To evaluate the effectiveness of different feature combinations on our model, we performed an ablation study across multiple datasets, analyzing the performance both within individual datasets and across datasets. Specifically, we examined our model on DFDC, WLDR, FF++ (face swap and expression swap), and NVFIAR datasets, testing each feature combination’s impact. Results for in-dataset evaluations are presented in Table VI, while cross-dataset results are provided in Table VII.

Starting with only deep identity features (DIF), the model established a baseline by achieving $88\pm 5\%$ AUC. With the addition of behavioral features, along with deep identity features, our model improves 9.65%, 2.95%, 5.20%, 4.90% and 6.25% AUC on DFDC, WLDR, FF-FS, FF-ES and NVFAIR, respectively. The AUC gains for geometric features, combined with deep identity and behavioral features, are 4.64%, 3.87%, 3.49%, 5.04% and 4.66%. Table VI clearly shows that the final feature set, with behavioral and geometric features, achieves a more stable performance on all datasets. For cross-dataset analysis, we trained the model on FF++ and evaluated it on DFDC, DFD, and CelebDF. For cross-dataset analysis, the model achieves an AUC of 73.90%, 76.52% and 75.54% on DFDC, DFD and CelebDF respectively, as shown in Table VII. The addition of behavioral features improves 12.90%, 10.86% and 10.89% AUC on DFDC, DFD and CelebDF. The AUC gains for final combination of features on cross-dataset evaluations are 5.25%, 4.24% and 3.29%. Our proposed combination of features achieves the best generalization towards new types of deepfakes and expressions swaps.