Deepfake detection in video data typically relies on mining deep embedded representations across multiple modalities to obtain discriminative fused features and thereby improve detection accuracy. However, existing approaches predominantly focus on how to exploit complementary information across modalities to ensure effective fusion, while often overlooking the impact of noise and interference present in the data. For instance, issues such as small objects, blurring, and occlusions in the visual modality can disrupt the semantic consistency of the fused features. To address this, we propose a Multimodality Semantic Consistency Fusion model for video forgery detection. The model introduces a semantic consistency gating mechanism to enhance the embedding of semantically aligned information across modalities, thereby improving the discriminability of the fused representations. Furthermore, we incorporate an event-level weakly supervised loss to strengthen the global semantic discrimination of the video data. Extensive experiments on standard video forgery detection benchmarks demonstrate the effectiveness of the proposed method, achieving superior performance in both forgery event detection and localization compared to state-of-the-art approaches.
1 January 2024 in The International Arab Journal of Information Technology
Semi-supervised learning is a powerful paradigm for excavating latent structures of between labeled and unlabeled samples under the view of models constructing. Currently, graph-based models solve the approximate matrix that directly represent distributions of samples by the spatial metric. The crux lies in optimizing connections of samples, which is achieved by subjecting to must-links or cannot-links. Unfortunately, to find links are rather difficult for semi-supervised clustering with very few labeled samples, therefore, significantly impairs the robustness and accuracy in such scenario. To address this problem, we propose the Cohesive Pair-wises Constrained deep Embedding model (CPCE) to obtain an optimal embedding for representing the category distribution of samples and avoid the failed graph-structure of the global samples. CPCE designs the deep network framework based on CNN-Autoencoder by minimizing reconstruct errors of samples, and build up constrains both of the sample distribution for within-class and the category distribution for intra-class to optimal the latent embedding. Then, leverage the strong supervised information obtained from cohesive pair-wises to pull samples into within-class, which avoid the similarity of high-dimension features located in different categories to achieve more the compact solution. We demonstrate the proposed method in popular datasets and compare the superiority with popular methods
Objective. Radiotherapy (RT) requires accurate and consistent patient positioning to ensure precise radiation delivery and minimize unnecessary exposure to healthy tissues. Conventional workflows rely heavily on clinicians’ experience and repeated CT-based registration, leading to inefficiency, patient discomfort, and potential alignment inconsistencies. This work aims to develop an automatic, robust, and low-cost posture offset detection method that overcomes these limitations. Approach. We propose a prompt-guided incremental fine-tuning model built upon a large-scale image segmentation backbone. The system captures real-time two-dimensional images from a single RGB camera and automatically generates adaptive prompt points based on individual body shapes and postures, improving segmentation robustness and reducing environmental interference. An incremental fine-tuning strategy enables continuous adaptation to newly collected patient images throughout the treatment cycle. Furthermore, a multi-level offset analysis framework is introduced, integrating contour-level, keypoint-level, and pixel-level estimations to identify, localize, and quantify posture deviations across multiple granularities. The system is deployed clinically to collect real RT data and construct a dedicated validation dataset. Main Results. Extensive experiments on real clinical data show that the proposed method achieves accurate, fast, and stable posture offset detection. It substantially improves positioning consistency and efficiency compared with conventional workflows. Ablation studies further demonstrate the effectiveness and necessity of each module within the framework. Significance. This study provides a practical and low-cost solution for RT positioning, reducing clinician workload and patient burden while improving treatment accuracy. It demonstrates the potential of prompt-guided incremental adaptation and multi-level offset analysis in real RT environments, offering a promising direction for intelligent, automated RT positioning systems.
Reliable condition monitoring in cyber-physical systems requires computational intelligence models that recover normal operating semantics from multivariate measurements that are noisy, non-stationary, and semantically sparse. This paper studies the mining of latent semantic correlations among variables from predominantly normal data. We propose StableAD-TS, a symmetric encoder–decoder that learns a noise-suppressed manifold of normal operation and reconstructs from that manifold, so that deviations appear as large residuals. In the encoder, spatiotemporal hybrid attention and Discriminative Semantic Gates (DS-Gates) extract hierarchical cross-variable features under semantic sparsity. In the decoder, stable-component skip connections and offset subtraction retain informative trends while isolating anomalous perturbations; a single-step diffusion perturbation further improves robustness under drift. Reconstruction residuals are used to assess event-level separability of the learned semantics. On four industrial monitoring benchmarks—PSM, SWaT, SMD, and MSL—StableAD-TS attains competitive or superior affiliation precision, recall, and F1 against 18 representative baselines, confirming the benefit of mining latent multivariate information.