With the advancement of intelligent driving and computer vision systems, high-quality and diverse rainy-day images are crucial for training robust visual systems. However, traditional image generation models rely on large-scale centralized datasets, which raises serious issues such as data privacy, security, and data silos when handling geographically distributed data. To address these challenges, this paper proposes a novel federated learning-based framework for collaboratively training a global flow matching model capable of generating high-quality rainy-day images with regional characteristics while preserving privacy. The framework utilizes the FedAvg algorithm to distribute the flow matching model to local servers for regional training, capturing unique geographical features. After training, the model parameters are securely aggregated to a central server under encryption protection, forming a powerful global model through weighted averaging. Through rigorous comparisons with the TPSeNCE framework, experimental results demonstrate that our method successfully achieves distributed collaborative model training without sacrificing image quality or model performance, offering a new approach for scalable generative modeling in distributed environments.

Owing to its rapid progress and broad application prospects, few-shot action recognition has attracted considerable interest. However, current methods are predominantly based on limited single-modal data, which does not fully exploit the potential of multimodal information. This paper presents a novel framework that actively identifies reliable modalities for each sample using task-specific contextual cues, thus significantly improving recognition performance. Our framework integrates an Active Sample Inference (ASI) module, which utilizes active inference to predict reliable modalities based on posterior distributions and subsequently organizes them accordingly. Unlike reinforcement learning, active inference replaces rewards with evidence-based preferences, making more stable predictions. Additionally, we introduce an active mutual distillation module that enhances the representation learning of less reliable modalities by transferring knowledge from more reliable ones. Adaptive multimodal inference is employed during the meta-test to assign higher weights to reliable modalities. Extensive experiments across multiple benchmarks demonstrate that our method significantly outperforms existing approaches.

Semantic communication is a transformative approach for efficient data transmission in bandwidth-constrained IoT networks, particularly for device-to-device (D2D) communication and real-time systems. To address the challenges of high-quality data reconstruction in low-bandwidth scenarios, such as IoT-enabled underwater or wireless networks, this article proposes a novel framework for ultraefficient image and video frame compression. The framework employs an optimized feature extraction method to capture essential semantic information from images or video frames, converting them into compact grayscale representations to minimize data volume for bandwidth-constrained devices. At the receiver, a dual-decoding strategy reconstructs structural details using lightweight semantic reconstruction techniques, followed by color attribute recovery, ensuring high visual quality in real-time applications. This approach enhances transmission efficiency, reduces latency, and maintains information integrity in resource-limited IoT environments. Experimental results on the Kodak dataset show a 7.02% compression ratio (PSNR = 13 dB), with a 9.76% improvement in bandwidth efficiency compared to existing methods. For 4K images and video frames, the framework achieves a 98.35% MS-SSIM retention rate under SNR conditions of 1–15 dB, demonstrating robust performance for real-time IoT communication.

Bitcoin, the first decentralized cryptocurrency system, uses a simple but effective difficulty adjustment algorithm to stabilize its average time of the block creation at 10 minutes. Over time, the volatility of the Bitcoin price has become higher and higher, and it causes the total hashrate (the hash power of the entire network) constantly fluctuating. Both facts and our experimental results prove that Bitcoin's difficulty adjustment algorithm cannot respond in time while the total hashrate is constantly fluctuating. Hence, we propose a consensus protocol with a real-time difficulty adjustment algorithm, RTPoW. RTPoW allows the blockchain to adjust the difficulty target of each block by predicting the real-time total hashrate, so the block time can remain stable even if the total hashrate is wildly fluctuating. To evaluate the effect of RTPoW, we implemented a simulator of an experimental environment and tested our algorithm. The results obtained have confirmed its effectiveness and stability.

With the rapid development of the Internet of Things (IoT), the continuous emergence of cyberattacks have brought great threat to the security of the network. Intrusion Detection System (IDS) which can identify malicious network attacks has become a strong tool to ensure network security. Many deep learning-based approaches have been used in IDS. However, most of these researches ignore the internal structural characteristics of the network traffic, and cannot accurately learn the key features of the malicious traffic. Thus, they have a low accuracy in classifying different kinds of network attacks. In this paper, we build an intrusion detection model DAL (Dense-Attention-LSTM, DAL), in which dense dilated convolutions is used to extract the underlying features of the network traffic. Then, attention mechanism is utilized to capture key features which represent the structural characteristics of traffic data. Moreover, CuDNN-based long short-term memory network is used to learn time-related information of the traffic while accelerating the convergence of the model. Finally, global maxpooling is adopted to compress data and to improve the generalization capabilities of the proposed model. Experimental results on UNSW-NB15 dataset show that the binary classification accuracy of the proposed model is up to 92.65%. Further, it can also identify various attacks with the accuracy of 81.28%. The performance of our model is better than some competing machine learning methods and some deep learning methods. We published our code at https://github.co-m/cKiNg37/IDS-model-DAL.