In cognitive Internet of Things (CIoT) systems, data incompleteness is a recurring issue caused by sensor malfunctions, hardware limitations, and unreliable network conditions. Traditional matrix completion methods provide a solution but often fail when faced with randomly missing entries or long consecutive gaps, resulting in distorted recovery outcomes. To address this, we present a context-aware (situational) matrix completion framework that integrates anomaly detection, probabilistic clustering, and predictive reasoning. The method begins by identifying and removing abnormal sensor values to prevent their propagation during reconstruction. The incomplete matrix is then clustered using a probabilistic approach, with each cluster assigned a plausibility score. The most plausible cluster forms the contextual foundation for recovery. A regression model based on relational pair extraction from contextual data predicts missing values. For entirely missing rows or columns, regression outputs are directly employed, while randomly scattered missing values are refined using maximum a posteriori (MAP) estimation. We evaluate the proposed method on a 21.25-year multi-sensor environmental dataset. Experimental results and cross-validation analysis demonstrate significant improvements in recovery accuracy (> 99.30
Public-key encryption with equality test ( ) is an important cryptographic primitive for cloud computing, as it enables equality testing over ciphertexts without revealing the underlying plaintexts. In multi-replica cloud environments, it is often desirable that equality-test results be publicly verifiable. To support this functionality, publicly verifiable ( ) has been introduced. However, existing schemes achieve only chosen-plaintext attack (CPA) security, which is insufficient for security-critical cloud applications. In this paper, we propose the first chosen-ciphertext secure scheme. In the proposed construction, authorized cloud servers collaboratively perform equality testing, while any party can publicly verify the correctness of the test result. Moreover, only the authorized cloud servers can generate a valid proof for the equality-test outcome. In the random oracle model, we prove that the proposed scheme achieves one-wayness under chosen-ciphertext attacks against authorized internal adversaries based on the Bilinear Diffie-Hellman assumption and the preimage resistance of the hash function, while simultaneously achieving indistinguishability under chosen-ciphertext attacks against unauthorized external adversaries based on the decisional Bilinear Diffie-Hellman assumption. We further prove that the correctness proof is existentially unforgeable under chosen-message attacks based on the Computational Diffie-Hellman assumption. Experimental results show that, compared with the latest scheme of Li et al. (TDSC 2024), our scheme strengthens the security level from CPA to CCA and reduces the computational cost of public verification to 0.76× that of Li et al.’s scheme, with only moderate additional overhead in the overall workflow.
With the rapid growth of the Internet of Things (IoT), mobile devices, autonomous robots, and wireless sensor networks, there is increasing demand for inertial navigation systems that provide high accuracy with low latency and power consumption. However, existing deep learning methods often suffer from high computational cost and limited generalization, restricting deployment on resource-constrained edge devices. To address these challenges, this paper proposes a lightweight Multi-Stage Extraction Network (MSENet) for inertial position estimation directly from raw inertial measurement unit (IMU) data. By employing a multi-stage feature extraction strategy, MSENet enhances temporal motion representation while reducing computational overhead. Experimental results on the RIDI and RONIN benchmarks show that MSENet achieves accuracy comparable to IMUNet while reducing parameters by 69.95
With the rapid advancement of Embodied Intelligence (EI), intelligent agents in wireless edge networks face the critical challenge of simultaneously maintaining visual perception quality and task inference reliability under bandwidth-constrained and noise-impaired wireless environments. Existing semantic communication systems usually prioritize either image reconstruction or task-oriented inference, making it difficult to support embodied perception tasks that require both perceptual recoverability and semantic understanding. To address this issue, this paper proposes an Enhanced Semantic Communication (E-SemCom) framework for embodied perception. The proposed framework integrates a heterogeneous dual-stream representation architecture with orthogonality-constrained feature learning to improve perceptual-semantic collaboration efficiency under limited communication resources. Furthermore, a serial reconstruction-guided collaborative inference strategy is developed to enhance downstream task robustness under adverse channel conditions. To comprehensively evaluate the collaborative perception capability, a Perceptual-Semantic-Inference (PSI) metric is adopted to jointly measure reconstruction fidelity and semantic effectiveness. Experimental results demonstrate that at a compression ratio of 1/12, the proposed method achieves superior overall PSI performance across a wide Signal-to-Noise Ratio (SNR) range from -5 dB to 20 dB. This work provides a novel framework for the efficient collaborative perception of embodied agents in dynamic wireless environments.
Traditional user authentication methods, based solely on passwords, are increasingly vulnerable to credential theft, creating a critical need for user-friendly and secure alternatives in Identity and Access Management (IAM) systems. In this paper, we propose Behavioral Two-Factor Authentication (BTFA), a novel user authentication model that enhances the traditional login process by integrating behavioral biometrics as a second layer of verification. By analyzing unique user-interaction patterns such as touch gestures, BTFA authenticates users without imposing any additional burdens. We utilize a Long Short-Term Memory (LSTM) network to analyze touch gestures and sensor data, serving as a dynamic second layer for user authentication during login. A field study was conducted using a custom-developed mobile app with registration and login interfaces, collecting 38542 behavioral samples for model training and validation. The model performance trained using sensor data achieved 97.28
