ABSTRACT Accurate and robust medical image analysis plays a critical role in disease screening, diagnosis, and prognosis. However, its development is often constrained by data scarcity, privacy concerns, and domain discrepancies. To address these challenges, we propose TransMedVision—a transitional training framework tailored for cross‐domain few‐shot medical image analysis tasks. The framework consists of three stages: (1) initializing the model with a vision backbone pretrained on large‐scale natural image datasets; (2) performing short‐term transitional training on intermediate medical image datasets to reduce the representation gap between natural and medical domains, while stabilizing feature learning; and (3) fine‐tuning on the target few‐shot CT dataset to obtain the final classifier. By preserving general visual features and gradually adapting them to medical domains, TransMedVision enhances both cross‐domain transfer accuracy and training stability. In cross‐domain few‐shot COVID‐19 pneumonia CT classification tasks, TransMedVision achieves state‐of‐the‐art performance (Accuracy = 0.9113, F1 = 0.9032, AUC = 0.9514). All datasets, code, and models are publicly released via https://github.com/01Matrix/TransMedVision to facilitate reproducibility and future research.

Human activity recognition (HAR) is an evolving technique that offers innovative solutions across various domains, such as healthcare, sports training, and human-computer interactions. This paper addresses the novel challenge of video-based activity recognition, focusing on detecting and classifying athletes' actions to enable precision sports training. Conventional HAR methods based on direct video analysis incur excessive computational overhead and constrained applicability. In contrast, our novel transformer-based framework, namely RSFomer, converts videos into multivariate time series, and then detects and classifies the athletes' actions. However, sports videos often suffer from severe occlusion, which introduces significant noise to the converted time series and thus deteriorates recognition performance. To address this challenge, we implement several innovative strategies to improve the robustness of our framework. First, we propose a dual-scale filtering mechanism that leverages the unscented Kalman filter and kinematic constraints to reduce noise and outliers in the converted time series. Second, we incorporate the masking mechanism and temporal slicing mechanism to enhance the transformer's ability to handle anomalies and extract multi-scale features for accurate action recognition. We perform extensive evaluations on our Boxing dataset as well as the UEA and FineGym datasets. The results demonstrate that our RSFomer is effective, outperforming existing state-of-the-art methods with significant advantages.

Accurately detecting out-of-scope queries is a challenging task in task-oriented dialog systems. Most existing research focus on adding an outlier detector after classification or designing an open world classification to identify unknown intents. There is still a major performance gap on achieving high efficiency and accuracy based on above methods. In our research, we tend to solve this problem by constructing an out-of-scope class in the classification. We propose an explainable centroid-based fine-tuning method including a modified decision metric (MDM) and a centroid-based cosine loss (CCL) on Pre-trained Transformer models for optimization. This loss function builds on Copernican structure and assigns the same margin to each in-scope class to resolve an ambiguous configuration on out-of-scope detection. Moreover, cosine similarity is utilized to remove radial variations of centroids. Experimental results show that our proposed method achieves improvement compared to other baseline methods.

Notary cross-chain transaction technologies have obtained broad affirmation from industry and academia as they can avoid data islands and enhance chain interoperability. However, the increased privacy concern in data sharing makes the participants hesitate to upload sensitive information without the trust foundation of the external network. To address this issue, this paper proposes a differential private notary mechanism (DPNM) to preserve privacy in blockchain interoperations. It establishes a fully trusted notary organization to conduct data perturbation before replying query to the external blockchain network. In addition, the DPNM contains two built-in privacy budget allocation schemes: Efficiency priority scheme (EPS) and Privacy priority scheme (PPS). These schemes unify the privacy preferences among different nodes based on multi-node consensus in the decentralized environment. The EPS can generate noise linearly and work efficiently, and the PPS reflects better on nodes’ preferences. This paper utilizes several metrics including mechanism errors, elapsed time, latency, and gas consumption to evaluate the performance of DPNM compared to the traditional mechanisms. The experiment results indicate that the proposed mechanism can meet privacy preferences among different nodes and provide better utility with little extra cost.

Blockchain has gradually attracted widespread attention from the research community of the IoT, due to its decentralization, consistency, and other attributes. It builds a secure and robust system by generating a backup locally for each participant node to collectively maintain the network. However, this feature brings some privacy concerns since all nodes can access the chain data, users’ sensitive information under risk of leakage. The local differential privacy (LDP) mechanism can be a promising way to address this issue as it implements data perturbation before uploading to the chain. While traditional LDP mechanisms cannot fit well with the blockchain since the requirements of a fixed input range, large data volume, and using the same privacy budget, which are practically difficult in a decentralized environment. To overcome these problems, we propose a novel LDP mechanism to split input numerical data and implement perturbation by digital bits, which does not require a fixed input range and large data volume. In addition, we use an iteration approach to adaptively allocate the privacy budget for different perturbation procedures that minimize the total deviation of perturbed data and increase the data utility. We employ mean estimation as the statistical utility metric under the same and randomized privacy budgets to evaluate the performance of our novel LDP mechanism. The experiment results indicate that the proposed LDP mechanism performs better in different scenarios, and our adaptive privacy budget allocation model can significantly reduce the deviation of the perturbation function to provide high data utility while maintaining privacy.