Diffusion-based methods have substantially improved the performance of full-body Text-to-Motion (T2M) generation from natural language descriptions. Despite this progress, accurately capturing the fine-grained semantics of composite prompts remains challenging. Approaches that rely solely on a single global text condition often fail to retain part-specific semantic cues, leading to deviations in the motions of certain body parts from the intended descriptions. Recent methods have attempted to address this by incorporating both global and local conditions, yet these are typically combined using fixed ratios or applied in separate stages, which restricts their adaptability to evolving semantic requirements during generation. To address these constraints, this work proposes the Embedding Gating Network (EGN), which dynamically modulates the contributions of global and local information according to the current noisy motion state and the diffusion timestep. By conditioning the gating mechanism on the intermediate noisy motion estimate, EGN adjusts the relative importance of global and local information to emphasize semantics that remain underrepresented at each denoising step. The conditioned signals are processed through independent part-wise generation pathways to minimize semantic interference, while a lightweight fusion module enables inter-part information exchange to preserve structural coherence across the full body. Experiments on the HumanML3D benchmark show that the proposed method consistently improves text-motion alignment over existing full-body and part-based baselines, without compromising motion quality or diversity. Analysis of the learned gating coefficients reveals that local conditions primarily contribute to the formation of part-wise structural outlines during early denoising stages, whereas global conditions become increasingly influential, integrating cross-part semantics and refining full-body consistency as denoising advances. These findings indicate that dynamically modulating conditioning signals during generation is an effective alternative to fixed-ratio conditioning.

Research has been conducted to reduce resource consumption in 3D medical image segmentation for diverse resource-constrained environments. However, decreasing the number of parameters to enhance computational efficiency can also lead to performance degradation. Moreover, these methods face challenges in balancing global and local features, increasing the risk of errors in multi-scale segmentation. This issue is particularly pronounced when segmenting small and complex structures within the human body. To address this problem, we propose a multi-stage hierarchical architecture composed of a detector and a segmentor. The detector extracts regions of interest (ROIs) in a 3D image, while the segmentor performs segmentation in the extracted ROI. Removing unnecessary areas in the detector allows the segmentation to be performed on a more compact input. The segmentor is designed with multiple stages, where each stage utilizes different input sizes. It implements a stage-skipping mechanism that deactivates certain stages using the initial input size. This approach minimizes unnecessary computations on segmenting the essential regions to reduce computational overhead. The proposed framework preserves segmentation performance while reducing resource consumption, enabling segmentation even in resource-constrained environments.

Federated Learning (FL), a practical solution that leverages distributed data across devices without the need for centralized data storage, which enables multiple participants to jointly train models while preserving data privacy and avoiding direct data sharing. Despite its privacy-preserving advantages, FL remains vulnerable to backdoor attacks, where malicious participants introduce backdoors into local models that are then propagated to the global model through the aggregation process. While existing differential privacy defenses have demonstrated effectiveness against backdoor attacks in FL, they often incur a significant degradation in the performance of the aggregated models on benign tasks. To address this limitation, we propose a novel backdoor defense mechanism based on differential privacy. Our approach first utilizes the inherent out-of-distribution characteristics of backdoor samples to identify and exclude malicious model updates that significantly deviate from benign models. By filtering out models that are clearly backdoor-infected before applying differential privacy, our method reduces the required noise level for differential privacy, thereby enhancing model robustness while preserving performance. Experimental evaluations on the CIFAR10 and FEMNIST datasets demonstrate that our method effectively limits the backdoor accuracy to below 15% across various backdoor scenarios while maintaining high main task accuracy.