Deepfake technology, which encompasses various video manipulation techniques implemented through deep learning algorithms-such as face swapping and expression alteration-has advanced to generate fake videos that are increasingly difficult for human observers to detect, posing significant threats to societal security. Existing methods for detecting deepfake videos aim to identify such manipulated content to effectively prevent the spread of misinformation. However, these methods often suffer from limited generalization capabilities, exhibiting poor performance when detecting fake videos outside of their training datasets. Moreover, research on the precise localization of manipulated regions within deepfake videos is limited, primarily due to the absence of datasets with fine-grained annotations that specify which regions have been manipulated.To address these challenges, this paper proposes a novel spatial-based training method that does not require fake samples to detect spatial manipulations in deepfake videos. By employing a technique that combines multi-part local displacement deformation and fusion, we generate more diverse deepfake feature data, enhancing the detection accuracy of specific manipulation methods while producing mixed-region labels to guide manipulation localization. We utilize the Swin-Unet model for manipulation localization detection, incorporating classification loss functions, local difference loss functions, and manipulation localization loss functions to effectively improve the precision of localization and detection.Experimental results demonstrate that the proposed spatial-based training method without fake samples effectively simulates the features present in real datasets. Our method achieves satisfactory detection accuracy on datasets such as FF++, Celeb-DF, and DFDC, while accurately localizing the manipulated regions. These findings indicate the effectiveness of the proposed self-blending method and model in deepfake video detection and manipulation localization.

In the context of cloud computing, network attackers usually exhibit complex, dynamic, and diverse behavior characteristics. Existing research methods, such as Bayesian attack graphs, lack evidence correlation and real-time reflection of the network attack events, and high computational complexity for attack analysis. To solve these problems, this study proposes a Dynamic Uncertain Causal Attack Graph (DUCAG) model and a Causal Chain-based Risk Probability Calculation (CCRP) algorithm. The DUCAG model is constructed to represent the uncertain underlying causalities among network attack events, and the CCRP algorithm aims at dynamically updating the causality weights among different network attack events and attacker hypotheses based on alarm information and causal chain reasoning process. By causality simplification and causality reasoning methods, the CCRP efficiently predicts the attacker behaviors and potential attack likelihood under uncertain time-varying attack situations, and is robust to the incompleteness and redundancy in alarm information. Four experiments under different attack scenarios demonstrate that, the DUCAG model can effectively characterize and predict the complex and uncertain attack causalities, in a manner of high time efficiency. The proposed method has application significance on cloud computing platforms by dynamically evaluating network security status, predicting the future behaviors of attackers, and assisting in adjusting network defense strategies.

In recent years, the research on text classification algorithm is still a hot topic in text mining. The KNN is a classic text classification algorithm. The rule of finding the nearest neighbors directly affects the performance and precision of categorization. In this paper, we mainly focus on distance measure and similarity. We propose a new text classification algorithm which combines KNN and Choquet integral. Choquet integral provides a new way to find the k-nearest neighbors. The result of experiment shows that the performance of this method is better than the classical distance measure or similarity measure.

The widespread dissemination of multimodal disinformation, which combines inflammatory text with manipulated images, poses a severe threat to society. Existing detection methods typically process textual and visual features in isolation or perform simple fusion, failing to capture the sophisticated semantic inconsistencies commonly found in false information. To address this, we propose a novel framework: Emotion-Topic Injection and Consistency Detection Network (ETICD-Net). First, a large language model (LLM) extracts structured sentiment and topic-guided signals from news texts to provide rich semantic clues. Second, unlike previous approaches, this guided signal is injected into the feature extraction processes of both modalities: it enhances text features from BERT and modulates image features from ResNet, thereby generating sentiment-topic-aware feature representations. Additionally, this paper introduces a hierarchical consistency fusion module that explicitly evaluates semantic coherence among these enhanced features. It employs cross-modal attention mechanisms, enabling text to query image regions relevant to its statements, and calculates explicit dissimilarity metrics to quantify inconsistencies. Extensive experiments on the Weibo and Twitter benchmark datasets demonstrate that ETICD-Net outperforms or matches state-of-the-art methods, achieving accuracy and F1 scores of 90.6% and 91.5%, respectively.

The advent of self-attention mechanisms within Transformer models has significantly propelled the advancement of deep learning algorithms, yielding outstanding achievements across diverse domains. Nonetheless, self-attention mechanisms falter when applied to datasets with intricate semantic content and extensive dependency structures. In response, this paper introduces a Diffusion Sampling and Label-Driven Co-attention Neural Network (DSLD), which adopts a diffusion sampling method to capture more comprehensive semantic information of the data. Additionally, the model leverages the joint correlation information of labels and data to introduce the computation of text representation, correcting semantic representation biases in the data, and increasing the accuracy of semantic representation. Ultimately, the model computes the corresponding classification results by synthesizing these rich data semantic representations. Experiments on seven benchmark datasets show that our proposed model achieves competitive results compared to state-of-the-art methods.