As AI-generated content (AIGC) thrives, deepfakes have expanded from single-modality falsification to cross-modal fake content creation, where either audio or visual components can be manipulated.While using two unimodal detectors can detect audio-visual deepfakes, cross-modal forgery clues could be overlooked.Existing multimodal deepfake detectors typically establish correspondence between the audio and visual modalities for binary real/fake classification and require the co-occurrence of both modalities.However, in real-world multi-modal applications, missing modality scenarios may occur where either modality is unavailable.In such cases, audio-visual detection methods are less practical than two independent unimodal methods.Consequently, the detector can not always obtain the number or type of manipulated modalities beforehand, necessitating a fake-modality-agnostic audio-visual detector.In this work, we introduce a comprehensive framework that is agnostic to fake modalities, which facilitates the identification of multimodal deepfakes and handles situations with missing modalities, regardless of the manipulations embedded in audio, video, or even cross-modal forms.To enhance the modeling of cross-modal forgery clues, we employ audio-visual speech recognition (AVSR)

The data-intensive nature of Diffusion models amplifies the risks of privacy infringements and copyright disputes, particularly when training on extensive unauthorized data scraped from the Internet. Membership Inference Attacks (MIA) aim to determine whether a data sample has been utilized by the target model during training, thereby serving as a pivotal tool for privacy preservation. Current MIA employs the prediction loss to distinguish between training member samples and non-members. These methods assume that, compared to non-members, members, having been encountered by the model during training result in a smaller prediction loss. However, this assumption proves ineffective in diffusion models due to the random noise sampled during the training process. Rather than estimating the loss, our approach examines this random noise and reformulate the MIA as a noise search problem, assuming that members are more feasible to find the noise used in the training process. We formulate this noise search process as an optimization problem and employ the fixed-point iteration to solve it. We analyze current MIA methods through the lens of the noise search framework and reveal that they rely on the first residual as the discriminative metric to differentiate members and non-members. Inspired by this observation, we introduce OMS, which augments existing MIA methods by iterating One More fixed-point Step to include a further residual, i.e., the second residual. We integrate our method into various MIA methods across different diffusion models. The experimental results validate the efficacy of our proposed approach.

Diffusion models pose risks of privacy breaches and copyright disputes, primarily stemming from the potential utilization of unauthorized data during the training phase. Membership inference is aimed to determine whether a specific sample has been used in the training process of a target model, representing a critical tool for privacy violation verification. However, the increased model complexity and stochasticity inherent in diffusion renders traditional shadow-model-based or metric-based methods ineffective when applied to diffusion models. Moreover, existing methods only yield binary classification labels which lack necessary comprehensibility in practical applications. In this paper, we explore a novel perspective for membership inference by leveraging the intrinsic generative priors within the diffusion model. Compared with unseen samples, training samples exhibit stronger generative priors within the diffusion model, enabling the successful reconstruction of substantially degraded training images. Consequently, we propose the Degrade Restore Compare (DRC) framework. In this framework, an image undergoes sequential degradation and restoration, and its membership is determined by comparing it with the restored counterpart. Experimental results verify that our approach not only significantly outperforms existing methods in terms of accuracy but also provides comprehensible decision criteria, offering evidence for potential privacy violations.

Recently, Deepfake has posed a significant threat to our digital society. This technology allows for the modification of facial identity, expression, and attributes in facial images and videos. The misuse of Deepfake can invade personal privacy, damage individuals’ reputations, and have serious consequences. To counter this threat, researchers have proposed active defense methods using adversarial perturbation to distort Deepfake products which can hinder the dissemination of false information. However, the existing methods are primarily based on image-specific approaches, which are inefficient for large-scale data. To address these issues, we propose an end-to-end approach to generate universal perturbations for combating Deepfake. To further cope with diverse Deepfakes, we introduce an adaptive balancing strategy to combat multiple models simultaneously. Specifically, for different scenarios, we propose two types of universal perturbations. Disrupting Universal Perturbation (DUP) leads Deepfake models to generate distorted outputs. In contrast, Lapsing Universal Perturbation (LUP) tries to make the output consistent with the original image, allowing the correct information to continue propagating. Experiments demonstrate the effectiveness and better generalization of our proposed perturbation compared with state-of-the-art methods. Consequently, our proposed method offers a powerful and efficient solution for combating Deepfake, which can help preserve personal privacy and prevent reputational damage.

With the rising prevalence of deepfakes, there is a growing interest in developing generalizable detection methods for various types of deepfakes. While effective in their specific modalities, traditional detection methods fall short in addressing the generalizability of detection across diverse cross-modal deepfakes. This paper aims to explicitly learn potential cross-modal correlation to enhance deepfake detection towards various generation scenarios. Our approach introduces a correlation distillation task, which models the inherent cross-modal correlation based on content information. This strategy helps to prevent the model from overfitting merely to audio-visual synchronization. Additionally, we present the Cross-Modal Deepfake Dataset (CMDFD), a comprehensive dataset with four generation methods to evaluate the detection of diverse cross-modal deepfakes. The experimental results on CMDFD and FakeAVCeleb datasets demonstrate the superior generalizability of our method over existing state-of-the-art methods. Our code and data can be found at https://github.com/ljj898/CMDFD-Dataset-and-Deepfake-Detection.

With the rapid advancements of large-scale text-to-image diffusion models, various practical applications have emerged, bringing significant convenience to society. However, model developers may misuse the unauthorized data to train diffusion models. These data are at risk of being memorized by the models, thus potentially violating citizens' privacy rights. Therefore, in order to judge whether a specific image is utilized as a member of a model's training set, Membership Inference Attack (MIA) is proposed to serve as a tool for privacy protection. Current MIA methods predominantly utilize pixel-wise comparisons as distinguishing clues, considering the pixel-level memorization characteristic of diffusion models. However, it is practically impossible for text-to-image models to memorize all the pixel-level information in massive training sets. Therefore, we move to the more advanced structure-level memorization. Observations on the diffusion process show that the structures of members are better preserved compared to those of nonmembers, indicating that diffusion models possess the capability to remember the structures of member images from training sets. Drawing on these insights, we propose a simple yet effective MIA method tailored for text-to-image diffusion models. Extensive experimental results validate the efficacy of our approach. Compared to current pixel-level baselines, our approach not only achieves state-of-the-art performance but also demonstrates remarkable robustness against various distortions.

One-shot talking head generation has no explicit head movement reference, thus it is difficult to generate talking heads with head motions. Some existing works only edit the mouth area and generate still talking heads, leading to unreal talking head performance. Other works construct one-to-one mapping between audio signal and head motion sequences, introducing ambiguity correspondences into the mapping since people can behave differently in head motions when speaking the same content. This unreasonable mapping form fails to model the diversity and produces either nearly static or even exaggerated head motions, which are unnatural and strange. Therefore, the one-shot talking head generation task is actually a one-to-many ill-posed problem and people present diverse head motions when speaking. Based on the above observation, we propose OSM-Net, aone-to-manyone-shot talking head generation network with natural head motions. OSM-Net constructs a motion space that contains rich and various clip-level head motion features. Each basis of the space represents a feature of meaningful head motion in a clip rather than just a frame, thus providing more coherent and natural motion changes in talking heads. The driving audio is mapped into the motion space, around which various motion features can be sampled within a reasonable range to achieve the one-to-many mapping. Besides, the landmark constraint and time window feature input improve the accurate expression feature extraction and video generation. Extensive experiments show that OSM-Net generates more natural realistic head motions under reasonable one-to-many mapping paradigm compared with other methods.

One-shot talking head generation produces lip-sync talking heads based on arbitrary audio and one source face. To guarantee the naturalness and realness, recent methods propose to achieve free pose control instead of simply editing mouth areas. However, existing methods do not preserve accurate identity of source face when generating head motions. To solve the identity mismatch problem and achieve high-quality free pose control, we present One-shot Pose-controllable Talking head generation network (OPT). Specifically, the Audio Feature Disentanglement Module separates content features from audios, eliminating the influence of speaker-specific information contained in arbitrary driving audios. Later, the mouth expression feature is extracted from the content feature and source face, during which the landmark loss is designed to enhance the accuracy of facial structure and identity preserving quality. Finally, to achieve free pose control, controllable head pose features from reference videos are fed into the Video Generator along with the expression feature and source face to generate new talking heads. Extensive quantitative and qualitative experimental results verify that OPT generates high-quality pose-controllable talking heads with no identity mismatch problem, outperforming previous SOTA methods.

Face reenactment aims to synthesis a photo-realistic video of the source face by imitating the motion and expression of the driving video while keeping the source appearance (i.e. identity). Although good results have achieved recently, most state-of-the-art methods remain vulnerable to extreme conditions, which greatly restricts the application in the real world. Among various extreme conditions, the large pose problem is the most common one. We clarify that the large pose problem is mainly caused by the severe motion change between the source image and the current driving frame. An intuitive solution is to divide the severe motion change into a sequence of subtle motions. Therefore, we propose a new scheme that exploring the temporal coherence between previous neighbor frame and current frame. The smaller motion change between consecutive frames help to solve the large pose problem. Furthermore, a calibration net is designed to eliminate the error accumulation of the previous step. Extensive experiments demonstrate that our method performs better on large pose face reenactment than the state-of-the-art in terms of large pose cases and visual quality.

One-shot talking head generation has received growing attention in recent years, with various creative and practical applications. An ideal natural and vivid generated talking head video should contain natural head pose changes. However, it is challenging to map head pose sequences from driving audio since there exists a natural gap between audio-visual modalities. In this work, we propose a Flow-guided One-shot model that achieves NaTural head motions(FONT) over generated talking heads. Specifically, we design a probabilistic CVAE-based model to predict head pose sequences from driving audio and source face. Then we develop a keypoint predictor that produces unsupervised keypoints describing the facial structure information from the source face, driving audio and pose sequences. Finally, a flow- guided occlusion-aware generator is employed to produce photo-realistic talking head videos from the estimated keypoints and source face. Extensive experimental results prove that FONT generates talking heads with natural head poses and synchronized mouth shapes, outperforming other compared methods.

Face-to-face communication is a common scenario including roles of speakers and listeners. Most existing research methods focus on producing speaker videos, while the generation of listener heads remains largely overlooked. Responsive listening head generation is an important task that aims to model face-to-face communication scenarios by generating a listener head video given a speaker video and a listener head image. An ideal generated responsive listening video should respond to the speaker with attitude or viewpoint expressing while maintaining diversity in interaction patterns and accuracy in listener identity information. To achieve this goal, we propose the Multi-Faceted Responsive Listening Head Generation Network (MFR-Net). Specifically, MFR-Net employs the probabilistic denoising diffusion model to predict diverse head pose and expression features. In order to perform multi-faceted response to the speaker video, while maintaining accurate listener identity preservation, we design the Feature Aggregation Module to boost listener identity features and fuse them with other speaker-related features. Finally, a renderer finetuned with identity consistency loss produces the final listening head videos. Our extensive experiments demonstrate that MFR-Net not only achieves multi-faceted responses in diversity and speaker identity information but also in attitude and viewpoint expression.

Image animation involves generating a video of a source image imitating the pose of a driving video. Despite recent advancements in the image animation task, most state-of-the-art methods remain vulnerable to large pose changes. In cases of large pose changes, existing methods struggle to model the complex nonlinear motion and yield distorted results, which greatly restricts their application in the real world. To tackle this problem, we present a novel approach called Chain-of-Pose (CoP) that decomposes large pose changes into a sequence of intermediate pose changes. This enables us to handle simplified pose changes and improves the accuracy of pose estimation. Furthermore, to better preserve the appearance of the source object, we introduce the Appearance Refinement Module (ARM) that effectively integrates the appearance texture feature of the source image with the structural pose feature from the pose chain. Our experimental results demonstrate that our method qualitatively and quantitatively outperforms state-of-the-art approaches on four diverse datasets, comprising talking faces, human bodies, and pixel animals. Notably, our approach significantly improves video quality in the case of large object pose changes. Our code is attached to the supplementary material.

Nowadays advanced facial manipulation techniques produce deepfake videos more realistically, which makes deepfake detection more difficult. To capture subtle and intricate artifacts, recent works attempt to enhance low-level textural information by attention-based framework. However, these methods require complex simulated data or extra supervision. Highly dependent on training settings, these methods not only have high training costs but also are prone to overfitting. To address this issue, we propose a novel perspective of deepfake detection via so-called prior-attention. Specifically, we introduce prior textural information, such as edge and noise, to model the attention maps explicitly. Benefiting from these natural “attention maps”, our model significantly enhances discriminative information without additional supervision. Furthermore, we design a Feature Abstraction Block (FAB) to facilitate cross-layer features interaction and insert it into distinct layers of CNN to detect the inconsistencies at multiple spatial levels. Extensive experiments demonstrate that our method achieves performance comparable to state-of-the-art methods.

Given a target face and a driving face, face reenactment aims to transfer attributes from the driving face to the target face. In the last decade, a great number of methods have been proposed to generate realistic reenacted faces. However, when these methods are applied to generate a smiling face, most of them can only get a mouth with blurry teeth, making the reenacted face unrealistic. This problem is mainly caused by incomplete tooth structure in the target face image under the setting of one-shot reenactment. In order to obtain smiling reenacted faces with detailed tooth structure, our method uses the tooth information from the driving face rather than the target face. Furthermore, to better represent the tooth structure and expressions of the driving face, we extract the texture with a carefully designed geometry-aware encoder. By training the encoder with tooth segmentation task and non-identity classification task, we acquire refined tooth representations and meanwhile derive the non-identity part of the driving face. We also design a specific generator to fuse the tooth texture features into the target face. Moreover, we add a mouth loss function to further ensure the high definition of the smiling reenacted face. We compare our method to existing state-of-the-art approaches. The experiments show that our method gets comparable results on non-smiling face reenactment and has superior performance on smiling face reenactment.

Face reenactment is a challenging task, as it is difficult to maintain accurate expression, pose and identity simultaneously. Most existing methods directly apply driving facial landmarks to reenact source faces and ignore the intrinsic gap between two identities, resulting in the identity mismatch issue. Besides, they neglect the entanglement of expression and pose features when encoding driving faces, leading to inaccurate expressions and visual artifacts on large-pose reenacted faces. To address these problems, we propose a Large-pose Identity-preserving face reenactment network, LI-Net. Specifically, the Landmark Transformer is adopted to adjust driving landmark images, which aims to narrow the identity gap between driving and source landmark images. Then the Face Rotation Module and the Expression Enhancing Generator decouple the transformed landmark image into pose and expression features, and reenact those attributes separately to generate identity-preserving faces with accurate expressions and poses. Both qualitative and quantitative experimental results demonstrate the superiority of our method.

Recently, more and more realistic facial manipulation images and videos, known as DeepFakes, have been created and rapidly circulated in social media. Therefore, it is crucial to develop effective and efficient methods to detect the malicious DeepFakes. Previous approaches all adopt a two-step pipeline with multiple separate models, i.e., first face detection and then face forensics, and lacks robustness against compressed data. In this paper, we propose an end-to-end framework for detection and localization of face manipulation, named DLFMNet, which effectively integrates face detection and face forensics into one model, avoiding intermediate processes like image cropping and feature re-extraction. In addition, to capture richer and more robust manipulated clues, we exploit multi-domain features that takes advantages of two different but complementary domains (i.e., RGB and noise). The evaluations on FaceForensics++ dataset demonstrate the effectiveness of our proposed DLFMNet. https://github.com/LightningChan/DLFMNet.

Recent advances in face generation and manipulation have enabled the creation of sophisticated face-swapped videos, also known as DeepFakes, which brings great potential threats to our society. Hence, it is crucial to develop effective approaches to distinguish them. Currently, face-swapped videos produced by existing methods are prone to exhibit some subtle spatial and temporal manipulated traces, which can be utilized as distinctive clues for face-swapped video detection. In this paper, we propose a unified framework, named FSSpotter, to explore rich spatial and temporal information in the video simultaneously. It consists of a Spatial Feature Extractor (SFE), which aims to discover spatial evidences within a single frame, and a Temporal Feature Aggregator (TFA), which is responsible for capturing temporal inconsistencies between frames. Moreover, a novel data processing strategy is adopted to highlight the inconsistencies of forged face with its surrounding regions. The evaluations on Deepfakes of FaceForensics++, DeepfakeTIMIT, UADFV and Celeb-DF datasets demonstrate that the proposed approach achieves better or comparable performance on AUC scores.

DeepFake videos are widely distributed on social media platforms, which has seriously affected the authenticity of digital media content, calling for robust DeepFake detection methods. Although numerous detection methods are formulated as frame-based binary classification, less attention has been paid to aggregate the features over individual frames to get a video-based judgement. We observed that for the detection of DeepFake videos, three different level forgery features from frame, clip and video can complement each other. We also found that discrete, large interval sampling strategy is more suitable for DeepFake detection, which can sample more complex video scenes, including multiple subjects, diverse facial expressions and head poses. In this work, we propose a hierarchical framework, using 2D convolutional neural networks for frame-level features extraction followed by a 1D convolutional aggregator to extract clip-level and video-level features, which can comprehensively exploit three different levels of features to make decisions. Evaluation was performed on four datasets, including DFDC, Celeb-DF, FaceForensics++ and UADFV, which provides competitive results compared to other methods. Experimental results of cross-test demonstrate that our hierarchical framework has excellent generalization performance in the face of unknown datasets.

Graph neural networks (GNNs), powerful deep representation learning methods for graph data, have been widely used in various tasks, such as recommendation systems and link prediction. Most existing GNNs are designed to learn node embeddings on homogeneous graphs. Heterogeneous information network (HIN) with various types of nodes and edges still faces great challenges for the heterogeneity and rich semantic information. To make full use of the heterogeneous information, many works try to manually design meta-paths, which are paths connected with two objects. They utilize meta-paths to capture more semantic information in heterogeneous graphs. However, manually designed meta-paths require domain knowledge and meta-path-based heterogeneous graph embedding methods only utilize the information of nodes with the same type, ignoring the impacts of the different types of nodes. We propose meta-path generation online for heterogeneous network embedding for all types of nodes, which can generate meta-paths and learn node embeddings simultaneously. Firstly, we exhaust all meta-paths within k-hop for specific nodes and apply a meta-path guided nodes aggregation. Secondly, we adopt an attention mechanism to select Top-N meta-paths with the largest attention coefficients for the semantic aggregation. The above two stages constitute one layer of our approach. Through stacking multi-layers, we can generate longer and more complex meta-paths. Without domain-specific preprocessing, extensive experiments on two datasets demonstrate that our proposed approach achieves better performance compared with other recent methods that require predefined meta-paths from domain knowledge.