The widespread use of mobile devices for all kinds of transactions makes necessary reliable and real-time identity authentication, leading to the adoption of face recognition (FR) via the cameras embedded in such devices. Progress of deep Convolutional Neural Networks (CNNs) has provided substantial advances in FR. Nonetheless, the size of state-of-the-art architectures is unsuitable for mobile deployment, since they often encompass hundreds of megabytes and millions of parameters. We address this by studying methods for deep network compression applied to FR. In particular, we apply network pruning based on Taylor scores, where less important filters are removed iteratively. The method is tested on three networks based on the small SqueezeNet (1.24M parameters) and the popular MobileNetv2 (3.5M) and ResNet50 (23.5M) architectures. These have been selected to showcase the method on CNNs with different complexities and sizes. We observe that a substantial percentage of filters can be removed with minimal performance loss. Also, filters with the highest amount of output channels tend to be removed first, suggesting that high-dimensional spaces within popular CNNs are over-dimensioned. The models of this paper are available at https://github.com/HalmstadUniversityBiometrics/CNN-pruning-for-face-recognition. © 2025.

Predicting drug–drug interactions (DDIs) is crucial for understanding and preventing adverse drug reactions (ADRs). However, most existing methods inadequately explore the interactive information between drugs in a self-supervised manner, limiting our comprehension of drug–drug associations. This paper introduces EDDINet: Enhancing Drug-Drug Interaction Prediction via Information Flow and Consensus-Constrained Multi-Graph Contrastive Learning for precise DDI prediction. We first present a cross-modal information-flow mechanism to integrate diverse drug features, enriching the structural insights conveyed by the drug feature vector. Next, we employ contrastive learning to filter various biological networks, enhancing the model's robustness. Additionally, we propose a consensus regularization framework that collaboratively trains multi-view models, producing high-quality drug representations. To unify drug representations derived from different biological information, we utilize an attention mechanism for DDI prediction. Extensive experiments demonstrate that EDDINet surpasses state-of-the-art unsupervised models and outperforms some supervised baseline models in DDI prediction tasks. Our approach shows significant advantages and holds promising potential for advancing DDI research and improving drug safety assessments. Our codes are available at: https://github.com/95LY/EDDINet_code. © 2024 The Authors

The AI-driven Internet of Things (AIoT) has been widely applied in the field of Internet of Vehicles (IoV) for vehicular cooperation. Federated Learning (FL), due to its ability to protect users' data privacy, reduce communication overhead, and facilitate real-time decision-making, is widely applied in the augmented intelligence of things for vehicles (AIoV). However, integrating FL with AIoV poses challenges, including the absence of fine-grained access control, insufficient safeguards for FL tasks and vehicle identities, inadequate security for data transmission, and shortcomings in protecting data storage. These vulnerabilities may lead to risks such as vehicle tracking, model information theft, and data tampering. To address these challenges, we propose a privacy preservation mechanism for AIoV via cloud-edge-vehicle hierarchical sharding blockchain. Firstly, we propose a hierarchical anonymous authentication scheme for IoV devices with stronger scalability and higher fault tolerance. Vehicles only know the attributes of each other or which shard they belong to. Secondly, we present a secure FL task assignment scheme for AIoV. Edge nodes utilize attribute-based encryption to deploy fine-grained FL tasks based on vehicle attributes. Only users who meet the attributes can decrypt the content, protecting FL tasks content and participant identities. Thirdly, we present a secure data transmission scheme between AIoV devices to protect the identity and data privacy of both parties, while also achieving non-interactive key agreement. Additionally, we propose a scalable secure data sharing and storage scheme based on hierarchical sharding blockchain, aiming to reduce storage overhead and minimize trust costs. © 2014 IEEE.

With the advent of the intelligent information processing era, more and more private sensitive data are being collected and analyzed for intelligent decision making tasks. Such information processing also brings many challenges with existing privacy protection algorithms. On the one hand, the algorithms based on data encryption compromise the integrity of the original data or incur high computational and communication costs to some extent. On the other hand, algorithms based on distributed learning require frequent sharing of parameters between different computing nodes, which poses risks of leaking local model information and reducing global learning efficiency. To mitigate the impact of these issues, a Quantum Fuzzy Federated Learning (QFFL) algorithm is proposed. In the QFFL algorithm, a Quantum Fuzzy Neural Network (QFNN) is designed at the local computing nodes, which enhances data generalization while preserving data integrity. In global model, QFFL makes predictions through the Quantum Federated Inference (QFI). QFI leads to a general framework for quantum federated learning on non-IID data with oneshot communication complexity, achieving privacy protection of local data and accelerating the global learning efficiency of the algorithm. The experiments are conducted on the COVID19 and MNIST datasets, and the results indicate that QFFL demonstrates superior performance compared to the baselines, manifesting in faster training efficiency, higher accuracy, and enhanced security. In addition, based on the fidelity experiments and related analysis under four common quantum noise channels, the results demonstrated that it has good robustness against quantum noises, proving its applicability and practicality. Our code is available at https://github.com/LASTsue/QFFL. © IEEE

Quantum-inspired neural networks (QNNs) have shown potential in capturing various non-classical phenomena in language understanding, e.g., the emgerent meaning of concept combinations, and represent a leap beyond conventional models in cognitive science. However, there are still two limitations in the existing QNNs: (1) Both storing and invoking the complex-valued embeddings may lead to prohibitively expensive costs in memory consumption and storage space. (2) The use of entangled states can fully capture certain non-classical phenomena, which are described by the tensor product with powerful compression ability. This approach shares many commonalities with the process of word formation from morphemes, but such connection has not been further exploited in the existing work. To mitigate these two limitations, we introduce a Quantum-inspired neural network with Hierarchical Entanglement Embedding (QHEE) based on finer-grained morphemes. Our model leverages the intra-word and inter-word entanglement embeddings to learn a multi-grained semantic representation. The intra-word entanglement embedding is employed to aggregate the constituent morphemes from multiple perspectives, while the inter-word entanglement embedding is utilized to combine different words based on unitary transformation to reveal their non-classical correlations. Both the number of morphemes and the dimensionality of the morpheme embedding vectors are far smaller than the counterparts of words, which would compress embedding parameters efficiently. Experimental results on four benchmark datasets of different downstream tasks show that our model outperforms strong quantum-inspired baselines in terms of effectiveness and compression ability. © 2024 Elsevier Ltd

Social media text can be semantically matched in different ways, viz paraphrase identification, answer selection, community question answering, and so on. The performance of the above semantic matching tasks depends largely on the ability of language modeling. Neural network based language models and probabilistic language models are two main streams of language modeling approaches. However, few prior work has managed to unify them in a single framework on the premise of preserving probabilistic features during the neural network learning process. Motivated by recent advances of quantum-inspired neural networks for text representation learning, we fill the gap by resorting to density matrices, a key concept describing a quantum state as well as a quantum probability distribution. The state and probability views of density matrices are mapped respectively to the neural and probabilistic aspects of language models. Concretizing this state-probability duality to the semantic matching task, we build a unified neural-probabilistic language model through a quantum-inspired neural network. Specifically, we take the state view to construct a density matrix representation of sentence, and exploit its probabilistic nature by extracting its main semantics, which form the basis of a legitimate quantum measurement. When matching two sentences, each sentence is measured against the main semantics of the other. Such a process is implemented in a neural structure, facilitating an end-to-end learning of parameters. The learned density matrix representation reflects an authentic probability distribution over the semantic space throughout the training process. Experiments show that our model significantly outperforms a wide range of prominent classical and quantum-inspired baselines. © 2024 Elsevier Ltd

Event argument extraction (EAE) is an important information extraction task, which aims to retrieve arguments from texts and classify them into predefined argument roles. In practice, argument roles are usually annotated with numerous types, which often arises schema-specific nature and long-tail type nature. Existing studies explore the role correlations without full utilization, which may limit the capability of argument extraction. This paper investigates two crucial correlations potentially benefiting to the task, namely intra- and inter-event role correlations. The intra-event role correlations consider the role dependency within an event, thus helping to capture event schema-specific nature. The inter-event role correlations leverage the relevance among roles across events, thus helping to learn beneficial features from other roles. To achieve the above ideas, we propose RCEAE from a new role correlation-enhanced perspective. Particularly, for intra-event role correlations, we devise a prompt-based cross encoder to capture role correlations from an event prompt, and retrieve arguments considering event schema information. For inter-event role correlations, we devise a multi-view graph-based role encoder to build relevance for role representations, accessing to beneficial features not only from their training data but also from their related roles across different event types. By incorporating both the correlation knowledge, we predict event arguments with a role-specific interactive decoder. We conduct experiments on three public benchmarks, ACE, RAMS and WIKIEVENTS. Empirical results show RCEAE achieves state-of-the-art F1 on all benchmarks and demonstrates the effectiveness of incorporating both intra- and inter-event role correlations. © 2025

There is growing pressure to manage and run urban rail transit networks as more people select this mode of transportation for their travel. It is, therefore, essential to create a precise system to predict passenger movements from origins to destinations (OD). The effectiveness of the present methods in simulating links between stations using real-time passenger flow data is limited. This paper suggests a Spatiotemporal Fusion Network for OD flow forecasting in urban rail travel. By examining previous OD data, this network creates spatiotemporal link graphs connecting stations that integrate spatiotemporal correlations of passenger flows. These graphs are incorporated into the forecasting system to forecast OD flows. Using actual passenger flow data from Shanghai and Hangzhou, we validate our MGLTN model and show almost 4% improvement in prediction accuracy (measured by MAE) over many state-of-the-art baseline models on both datasets. We also present a dwell score derived from anticipated train frequencies and passenger flows. Each station receives a rating based on this score, which represents its dwell qualities in comparison to predetermined norms. © 2025

In the context of smart aquaculture, real-time Multi-fish Tracking (MFT) poses a significant challenge. Existing Multi-Object Tracking (MOT) methods are often designed for objects with specific shapes or regular motion patterns, such as pedestrians or cars. The unique characteristics of fish, including their uniform appearance and deformation during motion, have been largely overlooked in research. To address this gap, we introduce the Uniform and Deformable Multi-fish Tracking (UD-MFT) benchmark. This dataset not only incorporates challenges related to uniform appearance and diverse deformable shapes of fish during motion in daily activities but also encompasses common MOT challenges like occlusion and disappearance. All sequences are sourced from industrialized aquaculture environments, providing a practical and relevant setting. To understand the distinctiveness of UD-MFT, we quantify the degrees of deformation, appearance, and occlusion levels within the dataset and compare them with tracking targets in existing datasets. Furthermore, to facilitate practical applications, we conduct a comprehensive evaluation of state-of-the-art real-time MOT models on UD-MFT, establishing a comparative baseline for accuracy and computational requirements. Additionally, we perform an in-depth analysis of the impact of deformation and appearance similarity on tracking accuracy. Finally, we provide reflections and recommendations concerning potential avenues for future research in this field. The proposed UD-MFT aims to serve as a robust platform for developing algorithms capable of handling fish with multiple motion patterns, thereby contributing to the advancement of intelligent fish farming. © 2024 Elsevier Ltd

With the emergence of the resource and equipment sharing concept, many enterprises and organizations begin to implement cross-domain sharing of devices, especially in the field of the Internet of Things (IoT). However, there are many problems in the cross-domain usage process of devices, such as access control, authentication, and privacy protection. In this paper, we make the following contributions. First, we propose a blockchain-based cross-domain authentication management system for IoT devices. The sensitive device information is stored in a Merkle tree structure where only the Merkle root is uploaded to the smart contract. Second, a detailed security and performance analysis is given. We prove that our system is secure against several potential security threats and satisfies validity and liveness. Compared to existing schemes, our schemes realize decentralization, privacy, scalability, fast off-chain authentication, and low on-chain storage. Third, we implement the system on Ethereum with varying parameters known as domain number, concurrent authentication request number, and Merkle tree leaf number. Experimental results show that our solution supports the management of millions of devices in a domain and can process more than 10,000 concurrent cross-domain authentication requests, consuming only 5531 ms. Meanwhile, the gas costs are shown to be acceptable. © IEEE