Search

Face forgery detection aims to identify manipulated or altered facial images or videos created using artificial intelligence. Existing detection methods exhibit favorable performance on high-quality videos, but the videos in daily applications are commonly compressed into low-quality formats via social media. The detection difficulty is increased by the poor quality, indistinct detail features, and noises such as artifacts in these images or videos. To address this challenge, we propose a multi-scale dual-branch network for compressed face forgery, called MDCF-Net, effectively capturing cross-domain forgery features at various scales in compressed facial images. The MDCF-Net comprises two branches: an RGB domain branch utilizing Transformers to extract multi-scale fine-texture features from the original RGB images; a frequency domain branch designed to capture artifacts in low-quality videos by extracting global spectral features as a supplementary measure. Then, we introduce a feature fusion module (FFM) based on multi-head attention to merge diverse feature representations in a spatial-frequency complementary manner. Extensive comparative experiments on public datasets such as FaceForensics++, Celeb-DF, and WildDeepfake demonstrate the significant advantage of MDCF-Net in detecting highly compressed and low-quality forged images or videos, especially in achieving state-of-the-art performance on the FaceForensics++ low-quality dataset. Our approach presents a new perspective and technology for low-quality face forgery detection.

How to achieve fast and accurate small object detection holds crucial theoretical and practical significance. However, this task encounters substantial challenges due to scale differences among instances in the scene, along with the scarcity of inherent features and weak representation of small instances. To alleviate the above problems, we propose a novel attention-based cascaded adaptive feature pyramid fusion network, CA2Det, to effectively improve the small object detection performance. First, to prevent the information degradation of small-instance features during training, we introduce the efficient and lightweight Shuffle Attention mechanism to highlight the features of small instances. Second, to mitigate the information conflicts arising from the scale inconsistency among instances, we design a double-layer cascaded adaptive fusion pyramid module, CAFP, which can effectively suppress the information conflicts while enabling full information exchange across layers. Finally, we combine sparse convolution to achieve efficient high-resolution input, providing richer geometric information of the instances. Compared to the baseline network, on the COCO benchmark dataset and the popular UAV dataset VisDrone, both of which contain a large number of small instances, the proposed method improves the detection accuracy mAP values by 1.1% and 2.2%, respectively, while having a good real-time detection speed.

Visual object detection is a computer vision technique based on deep learning. However, due to privacy issues, it is highly challenging to build an object detection model according to the current methods on the centrally stored training dataset. Federated learning is an effective approach to solving the challenge of training data collection by using distributed training. We propose FedVisionBC, a blockchain-based federated learning system for visual object detection that addresses the challenges of single point of failure, model poisoning attacks, and membership inference attacks in traditional federated learning. In the FedVisionBC system, we set up an aggregation node and a verification node instead of a central server to solve the single point of failure problem. We establish a security mechanism that uses encryption techniques, verification nodes, and smart contracts to resist model poisoning attacks. Experimental results show that FedVisionBC can accomplish the object detection task when the percentage of malicious clients is less than 60%. We also propose a new algorithm, ADPFedAvg, to prevent membership inference attacks, which relies on user-level differential privacy technology and the federated average algorithm. Experimental results indicate that ADPFedAvg can achieve a large-scale visual object detection model with differential privacy protection, while only a negligible cost in predictive accuracy.

At present, most of the perceptual hash methods for image copyright protection rely on manually designed feature extraction and mapping, whose detection accuracy is insufficient. Some schemes based on deep learning are designed to consider a limited variety of content retention operations, which are not enough to deal with the increasingly severe situation of image copyright protection. In response to this situation, a novel Convolutional Neural Network (CNN)-based perceptual image hashing scheme is introduced in this paper. In this scheme, the training images are classified according to their original images and a Hadamard matrix is used to generate a hash center for each class in Hamming space. Then the Convolution Neural Network learns the feature extraction process of the image automatically, constrained by central quantization and distinct quantization, so that the hash code of each image converges to the hash center of their class, and generates the final hash sequence. The proposed scheme can successfully strike a balance between perceived robustness and discrimination capacity. Based on the test results on large-scale test sets, $F_{1}$ scores, equal error rate (EER) and receiver operating characteristic (ROC) curves demonstrate the superiority of our scheme compared with some state-of-the-art schemes.