The goal of this work is to adapt Segment Anything Models (SAM) into crack segmentation tasks via automatic label generation, thus eliminating manual annotation cost. In this regard, an intuitive approach is to extract edges of crack samples and generate labels via the dilation and erosion processes for fine-tuning SAM. However, this simple solution cannot guarantee the quality of generated labels, as crack regions will be corrupted due to the imperfect edge detection. To this end, this paper proposes CoGenSAM, a novel Codebook-interactive Generative Labeling framework that enables an annotation-free SAM fine-tuning. To achieve this, in the first stage, we pre-train a vector-quantized variational auto-encoder (VQVAE) by reconstructing the synthesized crack-like structures for learning crack-aware priors within the codebook. In the second stage, these priors help another VQVAE serve as the restoration model to restore the randomly corrupted structures into uncorrupted ones. Specifically, we propose the crack-aware contrastive-interaction to maximize the mutual information with the above priors via codebook interaction. Then, high-quality labels can be generated by restoring corrupted labels from edge detection, contributing to an annotation-free SAM fine-tuning. We collect a new dataset, Bridge2025, to address the limited availability of related bridge-oriented benchmarks. Experiments show that our performance is close to fully-supervised methods.

The Industrial Internet of Things (IIoT) integrates cyber-physical systems, edge-cloud computing, Artificial Intelligence (AI), and advanced connectivity to enable intelligent, data-driven operations across manufacturing, energy, transportation, and healthcare. The convergence of Operational Technology (OT) and Information Technology (IT), together with pervasive AI-driven automation, complicates meeting the joint requirements of performance, adaptability, and security (PAS). This survey adopts an explicitly AI-centric view of PAS and reviews IIoT research spanning deep reinforcement learning (DRL), multi-agent reinforcement learning (MARL), digital twins (DTs), federated and collaborative learning, graph neural networks (GNNs), blockchain-assisted analytics, and lightweight and post-quantum cryptography. The reviewed approaches are analyzed across device, edge, fog, cloud, and application layers under industrial constraints such as latency, reliability, resource limits, safety requirements, and regulatory compliance. A PAS-centric taxonomy is introduced that links technique families to deployment layers and application domains, and recurring trade-offs among PAS dimensions are summarized. Open challenges are identified in multi-objective optimization, semantic interoperability, lifecycle validation, and governance, with directions outlined toward secure and sustainable Industry 5.0 IIoT ecosystems.

Automated crack recognition has achieved remarkable progress in the past decades as a critical task in structure health monitoring, to ensure safety and durability in many industrial scenarios. However, imbalanced crack recognition remains challenging due to the scarcity of crack samples and the consequential limited diversity. To resolve this, Artificial Intelligence Generated Content (AIGC) has been gradually adopted to generate synthetic data and reduce reliance on large amounts of labeled crack samples. This paper assumes that a crack sample in the feature space can be regarded as a combination of crack and background semantics. Then, the decompose-compose feature augmentation framework (DeCo) is proposed to perform crack data synthesis in the feature space by randomly composing crack and background semantic-relevant features. Specifically, the contrastive learning-based decomposing loss is proposed to enforce two encoders to separately learn crack and background semantics from crack samples with the theoretical guarantee. After that, an effective cross-instance feature union strategy is proposed to synthesize diverse crack samples by composing the crack-relevant features from a crack sample and background-relevant features across other training samples. To address the limited availability of related benchmarks, we collect INPP2022 and IRC2022 datasets from real-world applications in nuclear power plants and road pavement. Experimental results show that DeCo performs favorably against state-of-the-art competitors in imbalanced crack recognition tasks.

The widespread adoption of machine learning (ML) in privacy-sensitive domains such as healthcare and finance has amplified concerns surrounding data privacy, regulatory compliance, and generalization across data domains. Federated learning (FL) offers partial mitigation by enabling collaborative training without sharing raw data. However, current FL frameworks often fail to address key challenges, including vulnerabilities to semi-honest and colluding participants, inefficiencies in computation and communication for edge devices, and limited support for uncertainty-aware models like Gaussian Processes (GPs). This paper presentsF2GP, a secure federated framework for sparse GP approximation that addresses these gaps through two core innovations. First, a privacy-preserving support set selection algorithm based on one-shot federated$k$-means clustering improves model accuracy with minimal communication overhead. Second, a multi-party homomorphic encryption (HE) scheme secures the aggregation and inference process, ensuring robustness even when up to$M-1$participants collude with the server. We implement F2GP on a testbed of 15 NVIDIA Jetson Nano devices, emulating low-resource edge environments. The framework demonstrates practical scalability and achieves 15–20% better accuracy compared to non-secure distributed and centralized GP baselines, while maintaining sub-linear communication overhead. Evaluations across diverse datasets confirm F2GP's performance, resilient privacy, and suitability for uncertainty-aware inference. The code will be made available onhttps://gitee.com/SZU-AI4H/securedgp

Edge computing improves the performance of Internet of Things (IoT) devices by moving cloud services closer to where these devices operate, thereby reducing delays and saving bandwidth. However, the IoT devices communicate wirelessly with edge servers, which creates a significant challenge in keeping the devices' locations private, especially when they use different methods. To address this, researchers have proposed various methods. However, these methods require a lot of computational power and storage, which makes them unsuitable for edge computing environments. To address these challenges, we propose a certificateless cryptography (CLC) and identity-based cryptography (IBC) based heterogeneous deniable authentication (CI-HDA) scheme. It enables an IoT device operating CLC to securely communicate with an edge server using IBC. The server verifies the devices authenticity without proving its participation to third parties, which ensures location privacy in edge computing environments. Additionally, our scheme supports batch verification, which enables the server to efficiently validate multiple authenticators simultaneously. The security of CI-HDA scheme is formally proven in the random oracle model, and performance evaluations demonstrate reduced computational and communication/storage overheads compared to existing methods. We also explored its application in military surveillance, which highlights its practicality in privacy-sensitive edge computing environments.

Intelligent Transportation Systems (ITS) rely on the Internet of Vehicles (IoV) to sustain high data rates and low latency under dynamic and heterogeneous conditions. Joint power and spectrum control across macro and micro tiers remains challenging due to mobility, interference coupling, and large continuous action spaces. EdgeSAC is a graph-aware Soft Actor Critic (SAC) framework executed at the edge for power control in hierarchical Fifth-Generation New Radio (5G NR) Multiple-Input Multiple-Output (MIMO) networks. A permutation-equivariant Graph Neural Network (GNN) with edge updates encodes co-channel interference among Base Stations (BSs) and outputs node-level power fractions under tier budgets. An on-demand scheduler activates fixed-size channels and assigns at most one macro and one micro resource per user to realize dual connectivity. Signal-to-Interference-plus-Noise Ratio (SINR) is mapped to rate using a Shannon with gap model with rank adaptive MIMO, enabling tier aggregation without action discretization. In simulation with Third Generation Partnership Project (3GPP) TR 38.901 path loss and Manhattan mobility, EdgeSAC increases throughput over SAC and Proximal Policy Optimization (PPO) and reduces power relative to Twin Delayed Deep Deterministic Policy Gradient (TD3), which raises energy efficiency and fairness. The findings indicate that interference-aware graph embeddings combined with entropy regularized continuous control provide a scalable and power-efficient solution for hierarchical IoV resource management.

Acute lymphoblastic leukemia is a childhood cancer prevalent worldwide, which can prove fatal within weeks or months. However, current diagnosis models based on machine learning and deep learning methods fail to consider device noise (pixel-level perturbations) and rotation/translation (spatial-transformed perturbations), which can undermine the model's robustness. Adversarial training is a potential solution to this issue. This paper presents a hybrid perturbation adversarial training (HPAT) strategy that leverages two types of adversarial samples: pixel-level adversarial samples and spatial adversarial samples. This work generates these hybrid adversarial samples through Projected Gradient Descent (PGD) in couple with spatial transformation based on the Bayesian optimization (STBO) algorithm, respectively. This work introduced the Mixed Batch Normalization (MixBN) module to handle both adversarial samples and clean samples, alleviating the problem of clean accuracy degradation due to adversarial training. The proposed hybrid adversarial training strategy is tested on the public acute lymphoblastic leukemia dataset and found that it outperformed existing acute lymphoblastic cell classification models.

Deep neural networks (DNNs) have substantially achieved high predictive accuracy in many vision tasks. However, we find that they are poorly calibrated for crack recognition tasks, as these DNNs tend to produce both under-confident and over-confident predictions in such safety-critical applications, thereby limiting their practical use in real-world scenarios. To address this issue, we propose a novel attack-inspired calibration loss (AICL) that explicitly regularizes class probabilities to be better confidence estimation. Specifically, we first propose the attack-inspired correctness estimation method (ACE) that aims to estimate the correctness degree of each sample via adversarial attacks. Then, we propose Correctness-aware Distribution Guidance, which starts from a distribution perspective that enforces the ordinal ranking of the predicted confidence referring to the estimated correctness degree. The proposed method can be conveniently implemented on top of any DNNs-based crack recognition model by serving as a plug-and-play loss function. To address the limited availability of related benchmarks, we collect a fully annotated dataset, namely, Bridge2024, which involves inconsistent cracks and noisy backgrounds in real-world bridges. Our AICL outperforms the state-of-art calibration methods on various benchmark datasets including CRACK2019, SDNET2018, and our BRIDGE2024.

Edge computing brings cloud services closer to Internet of Things (IoT) devices by processing data at the edge of the network, reducing latency and bandwidth usage. However, preserving the location privacy of IoT devices is a major challenge due to the heterogeneous, open wireless, and dynamic nature of the edge computing environments. The authentication process in this context may potentially disclose the device’s location, as IoT devices and edge server often rely on different security mechanisms. To address this concern, researchers have proposed various schemes to preserve location privacy. However, these schemes often incur high computational and communication overhead, making them unsuitable for IoT devices, which have limited processing power and storage capacity. In response to this challenge, we present a heterogeneous deniable authentication scheme based on certificateless cryptography (CLC) and public-key infrastructure (PKI), abbreviated as CP-HDA. This scheme enables an IoT device in a CLC to securely transmit messages to an edge server in a PKI. Using the CP-HDA scheme, the edge server can verify the legitimacy of the IoT device but cannot provide evidence to third parties regarding the device’s involvement in communication, thereby preserving the location privacy of the IoT device. It also supports batch verification, which enables faster verification of multiple deniable authenticators. The security of the CP-HDA scheme is formally proven in the random oracle model based on the assumption that the elliptic curve computational Diffie-Hellman problem is hard. In comparison to alternative schemes, our scheme enhances performance by reducing both computational and communication costs.

Homomorphic encryption (HE) is one of the more sophisticated methods of homomorphic cryptography (HC). HC efficiently contacts the interacting parties in open IoT and light-weighted network environments. This approach is capable of analyzing encrypted data without decryption. The operations use private and public keys. Then, during the assessment or evaluation, users may access the original data. Before conducting tests or evaluations, the customer must first encrypt the data and then decrypt it. Since consumers use several main cycles for the whole operation, which creates noise and computation overheads, the growth rate of computation overheads has increased. The growing ratio of noise to computation rate can interrupt the whole system, resulting in machine instability, protection, and privacy concerns. To resolve the security and privacy issues, the proposed schemes used different hardness assumptions, such as over-integer, learning with error, ideal lattices, bootstrapping, etc. In this article, we presents a comprehensive review of HE and its many varieties. The numerous possible applications of HE are covered at a high level in order to highlight the extent to which HE is used in the IoT and other lighted-weighted intelligent industry environments in a variety of various domains.

With the expanding use of Unmanned Aerial Vehicles (UAVs) in a variety of applications, their safety has become a critical concern. UAVs are confronted with a variety of unforeseen circumstances during missions; in these instances, the UAVs need to autonomously locate a suitable landing site, plan flight routes, and avoid obstacles in unstructured environments. Due to the limitations of computing power and sensors, it is challenging to attain the goal. The aim of research study is to investigate the monocular emergency autonomous landing algorithm. This work concentrates specifically on extracting depth and vision information. A topology information extractor is designed to transform images into graphs and assess the connectivity of terrain. In addition, a depth information extractor is designed to compute the slope and roughness of the ground. A$3$D topology optimizer is designed to optimize the graph based on depth information and evaluate the landing suitability using a heuristic strategy. For action decision making, a 3D topology decision method based on Depth-Enhanced Graph Structure (DEGS) is proposed. In order to demonstrate the efficacy of DEGS, this study constructed a simulated scenario based on an actual scene. The results of the experiment indicate that DEGS outperforms its counterparts in terms of the accuracy of action prediction and landing success rate.Note to Practitioners—Using DEGS, this study proposes a novel method for emergency UAV landing on unknown fields. The proposed method is founded on the following fundamental concepts: First, the UAV first generates a DEGS of the unknown field. Second, the UAV then evaluates the landing risk and guides the UAVs in planning a safe landing trajectory. Third, the UAV implements the landing trajectory and lands safely on the unknown field using monocular aerial vision. Frame Sequential and Self-supervision Network (FSSN), a multi-scale vision-based Monocular Depth Estimation (MDE) network, is proposed to estimate the depth map for real-time UAV flight phases. This method has been evaluated using simulations and real-world dataset of simulation images of monocular continuous frames and has shown to be effective in landing UAVs on unknown fields. A human-in-the-loop learning approach is proposed for updating DEGS with dynamic terrain classification that made the procedure feasible for unknown fields. In terms of landing success rate and action prediction accuracy, the results demonstrate that the proposed method is capable of landing a UAV on unknown terrain in a safe and efficient manner, and it is particularly useful in emergency situations where the UAV does not have prior knowledge of the field.

Crack segmentation receives extensive attention in structure health monitoring for many industrial scenarios, e.g., bridges, highways, and nuclear power plants. The current deep learning-based crack segmentation models enjoy the ability to extract discriminative crack features by training with an extensive labeled crack dataset. However, collecting extensive crack samples with accurate annotations from experts for a new scenario is labor-intensive, thereby limiting the effectiveness of these deep models in practical applications. To address this problem, the existing Fourier-based augmentation adopts a vanilla amplitude fusion process, i.e., the portion of amplitude components is fixed or randomly selected, failing to guarantee augmented samples’ semantics consistency, and diversity concerning the original sample. To fill this, this article proposes a self-adaptive Fourier augmentation framework that efficiently synthesizes diverse crack samples for training crack segmentation models. Our proposed framework advances Fourier transformation in an adversarial learning manner, alternating between self-adaptive Fourier-based data augmentation and teacher–student learning. The former aims to guarantee the diversity and semantics consistency of Fourier-based augmented samples, while the latter progressively updates the student network by observing these augmented samples for extracting discriminative features via a knowledge distillation mechanism. It is worth noting that the proposed method is only applied in the training stage without extra computation and memory during inference. Extensive experiments demonstrate the superiority of our method over the existing methods.

Tumor heterogeneity presents a significant challenge in predicting drug responses, especially as missense mutations within the same gene can lead to varied outcomes such as drug resistance, enhanced sensitivity, or therapeutic ineffectiveness. These complex relationships highlight the need for advanced analytical approaches in oncology. Due to their powerful ability to handle heterogeneous data, graph convolutional networks (GCNs) represent a promising approach for predicting drug responses. However, simple bipartite graphs cannot accurately capture the complex relationships involved in missense mutation and drug response. Furthermore, Deep learning models for drug response are often considered "black boxes", and their interpretability remains a widely discussed issue. To address these challenges, we propose an Interpretable Dynamic Directed Graph Convolutional Network (IDDGCN) framework, which incorporates four key features: 1) the use of directed graphs to differentiate between sensitivity and resistance relationships, 2) the dynamic updating of node weights based on node-specific interactions, 3) the exploration of associations between different mutations within the same gene and drug response, and 4) the enhancement of interpretability models through the integration of a weighted mechanism that accounts for the biological significance, alongside a ground truth construction method to evaluate prediction transparency. The experimental results demonstrate that IDDGCN outperforms existing state-of-the-art models, exhibiting excellent predictive power. Both qualitative and quantitative evaluations of its interpretability further highlight its ability to explain predictions, offering a fresh perspective for precision oncology and targeted drug development.

Distributed sparse Gaussian process (dGP) models provide an ability to achieve accurate predictive performance using data from multiple devices in a time efficient and scalable manner. The distributed computation of model, however, risks exposure of privately owned data to public manipulation. In this paper we propose a secure solution for dGP regression models using multi-key homomorphic encryption. Experimental results show that with a little sacrifice in terms of time complexity, we achieve a secure dGP model without deteriorating the predictive performance compared to traditional non-secure dGP models. We also present a practical implementation of the proposed model using several Nvidia Jetson Nano Developer Kit modules to simulate a real-world scenario. Thus, secure dGP model plugs the data security issues of dGP and provide a secure and trustworthy solution for multiple devices to use privately owned data for model computation in a distributed environment availing speed, scalability and robustness of dGP.

The Machine-Learning-as-a-Service (MLaaS) framework allows one to grab low-hanging fruit of machine learning techniques and data science, without either much expertise for this sophisticated sphere or provision of specific infrastructures. However, the requirement of revealing all training data to the service provider raises new concerns in terms of privacy leakage, storage consumption, efficiency, bandwidth, etc. In this paper, we propose a lightweight privacy-preserving MLaaS framework by combining Compressive Sensing (CS) and Generative Networks. It’s constructed on the favorable facts observed in recent works that general inference tasks could be fulfilled with generative networks and classifier trained on compressed measurements, since the generator could model the data distribution and capture discriminative information which are useful for classification. To improve the performance of the MLaaS framework, the supervised generative models of the server are trained and optimized with prior knowledge provided by the client. In order to prevent the service provider from recovering the original data as well as identifying the queried results, a noise-addition mechanism is designed and adopted into the compressed data domain. Empirical results confirmed its performance superiority in accuracy and resource consumption against the state-of-the-art privacy preserving MLaaS frameworks.

Deep Neural Networks (DNNs) have demonstrated remarkable performance in classification and regression tasks on RGB-based pathological inputs. The network’s prediction mechanism must be interpretable to establish trust in the clinical routine. One principal approach to interpretation is feature attribution. Feature attribution methods identify the importance of input features for the output prediction. Building on the Information Bottleneck Attribution (IBA) method, we recognize the RGB’s input regions with high mutual information with the network’s output for each prediction. IBA identifies input regions that have sufficient predictive information. In this paper, we introduce "IBATree", a novel approach that combines IBA with decision trees to enhance both the interpretability and accuracy of DNNs in cancer cell classification. Our method leverages the information bottleneck framework to inject noise into feature maps and then isolates the most informative features for model predictions while maintaining high classification performance. We evaluated our proposed approach on three datasets—CNMC, ISBI2016, and BreaKHis—demonstrating competitive accuracy and producing clear interpretations. This makes IBATree particularly suitable for clinical applications, where understanding the rationale behind predictions is crucial. Our results show that IBATree provides reliable predictions and valuable insights into feature importance, paving the way for its application in various biomedical domains.

The challenge of interpretability remains a significant barrier to adopting deep neural networks in healthcare domains. Although tree regularization aims to align a deep neural network’s decisions with a single axis-aligned decision tree, however, relying on one tree for all inputs often leads to sub-optimal performance and interoperability. To address this limitation, we propose an enhanced tree regularization method that integrates a post-hoc visual explainable model such as Grad-CAM. This approach guides the deep model to be well-approximated by decision trees tailored to the salient regions identified by Grad-CAM in the input space. We rigorously validate the effectiveness of this framework on two cancer cell datasets: CNMC, which focuses on acute lymphoblastic leukemia cells, and ISBI2016, which comprises benign and malignant skin lesions. The results demonstrate that the proposed method delivers simpler and more interpretable explanations without compromising accuracy, thereby advancing the interpretability of deep learning models in critical healthcare applications.

Crack segmentation datasets make great efforts to ob-tain the ground truth crack or non-crack labels as clearly as possible. However, it can be observed that ambiguities are still inevitable when considering the marginal non-crack re-gion, due to low contrast and heterogeneous texture. To solve this problem, we propose a novel clustering-inspired representation learning framework, which contains a two-phase strategy for automatic crack segmentation. In the first phase, a pre-process is proposed to localize the marginal non-crack region. Then, we propose an ambiguity-aware segmentation loss (Aseg Loss) that enables crack segmentation models to capture ambiguities in the above regions via learning segmentation variance, which allows us to further localize ambiguous regions. In the second phase, to learn the discriminative features of the above regions, we propose a clustering-inspired loss (CI Loss) that alters the supervision learning of these regions into an unsupervised clus-tering manner. We demonstrate that the proposed method could surpass the existing crack segmentation models on various datasets and our constructed CrackSeg5k dataset.

With recent progress in Internet of Things technology, it is becoming more and more commonplace to use unmanned aerial vehicles (UAVs) for inconsiderable purposes. On the other hand, traditional all of these UAV networks adopt a susceptible open wireless communication, rendering these systems vulnerable to attacks like eavesdropping, tampering, interrupting, and forging. The most effective way to address these security challenges is through signcryption. However, current signcryption methods are computationally and bandwidth-intensive, making them unsuitable for UAVs with limited resources and ground stations (GS) handling a high volume of messages. To address these challenges, we propose a solution employing an identity-based online/offline signcryption scheme to secure communication from a UAV to GS, known as IOOSC-U2G. This scheme leverages elliptic curve cryptography without the need for time-intensive operations like bilinear pairing. During the online phase, the absence of point multiplication operations, already executed in the offline phase, significantly alleviates computational burdens. This optimization significantly reduces computational overhead throughout the entire signcryption process of messages. Moreover, the IOOSC-U2G scheme ensures the privacy of UAV identities during communication with GS. Additionally, the proposed scheme empowers the GS to verify multiple inputs at once through batch verification method. We demonstrate that within the random oracle model, the IOOSC-U2G scheme guarantees security, specifically confidentiality and unforgeability, relying on the computational hardness assumptions of the elliptic curve inverse Computational Diffie-Hellman problem and the elliptic Discrete Logarithm problem, respectively. Moreover, our scheme outperforms current methods, particularly in computational and communication efficiency.

Protein methylation is one of the most important reversible post-translational modifications (PTMs), playing a vital role in the regulation of gene expression. Protein methylation sites serve as biomarkers in cardiovascular and pulmonary diseases, influencing various aspects of normal cell biology and pathogenesis. Nonetheless, the majority of existing computational methods for predictingprotein methylation sites(PMSP) have been constructed based on protein sequences, with few methods leveraging the topological information of proteins. To address this issue, we propose an innovative framework for predicting Methylation Sites using Graphs (GraphMethySite) that employs graph convolution network in conjunction with Bayesian Optimization (BO) to automatically discover the graphical structure surrounding a candidate site and improve the predictive accuracy. In order to extract the most optimal subgraphs associated with methylation sites, we extend GraphMethySite by coupling it with a hybrid Bayesian optimization (together named GraphMethySite$^+$) to determine and visualize the topological relevance among amino-acid residues. We evaluated our framework on two extended protein methylation datasets, and empirical results demonstrate that it outperforms existing state-of-the-art methylation prediction methods.

Crack detection has attracted extensive attention in an intelligent transportation system (ITS). Despite the substantial progress of deep learning technology on crack recognition tasks, due to the various limitations in traffic, equipment, and time, it is hard to collect copious samples for training deep models. Considering this, implicitly semantic data augmentation (ISDA) tries to augment the training set in the feature space. However, when applying it to crack recognition tasks, our empirical studies reveal that those poor-classified augmented samples have little semantic relevance to the crack class, resulting in a non-negligible negative effect on training deep models. Since the augmented features follow the multivariate normal distribution, it is computationally inefficient to explicitly sample those features and filter out the hard-classified augmented features. To this end, we propose the implicit gradient-modulated semantic data augmentation (IGMSDA) for addressing the above problems. Concretely, this paper first proposes gradient-modulated (GM) loss to dynamically modulate the gradient of those poor-classified augmented samples by reshaping the standard cross-entropy loss. And then, in the feature space, we derive an upper bound of the expected GM loss on the augmented training set to avoid the costly explicit sampling process. Experiments show that IGMSDA improves the generalization performance of the existing deep models on crack recognition datasets.

The annotation of Open Reading Frames (ORFs) is a crucial step in gene annotation, as it precisely delineates the specific regions of expressed genes. However, small Open Reading Frames (smORFs), in comparison to ORFs, are shorter in length, exhibit lower expression abundance, and are more challenging to predict. Particularly in the presence of noise in prokaryotic data and limited availability of positive sample data, the difficulty of prediction is amplified. Therefore, it is necessary to study smORF prediction methods. However, current machine learning models use limited data for modeling and overlook the existence of undiscovered positive samples within the negative samples. Additionally, they do not incorporate prior knowledge that can be calibrated to enhance the 3-nt periodicity. This work utilizes a multimodal VAE for data dimensionality reduction and employs a GAN to generate latent vectors for data augmentation. It incorporates PU learning to leverage unknown samples and combines Riboseq data from experiments with and without antibiotic treatment. Additionally, an adversarial training mechanism is employed to enhance the model’s robustness.

Retinopathy of Prematurity (ROP) is a retina disorder that affects premature infants with lower weights. If the patient cannot get the treatment in time when the illness reaches the last stage, irreversible vision loss will be caused. Nevertheless, there has been relatively little consideration given to the segmentation of the ridge, the key clinical characteristic of the illness. Additionally, existing research has not adequately addressed several segmentation issues, such as fragmentary topology, class imbalance, and false positives. This paper proposes a Densely Dilated U-Net (DD-UNet) improved from U-Net to tackle these challenges. Furthermore, the post-processing techniques based on the spatial relationship between vessels and ridges, along with the relative pixel counts of ridges and false positive results is integrated to mitigate false positive results in the predicted ridge. To enhance the precision of thin vessel, a sliding window sampling method is introduced for refined training. Compared with the state-of-the-art models in medical image segmentation, DD-UNet performs well in curvilinear structure segmentation of fundus image. For instance, our DD-UNet outperforms the Attention U-Net by 6.26% in terms of sensitivity and exhibits a 1.85% higher dice score in ridge segmentation.

Cracks are usually curve-like structures that are the focus of many computer-vision applications (e.g., road safety inspection and surface inspection of the industrial facilities). The existing pixel-based crack segmentation methods rely on time-consuming and costly pixel-level annotations. And the object-based crack detection methods exploit the horizontal box to detect the crack without considering crack orientation, resulting in scale variation and intra-class variation. Considering this, we provide a new perspective for crack detection that models the cracks as a series of sub-cracks with the corresponding orientation. However, the vanilla adaptation of the existing oriented object detection methods to the crack detection tasks will result in limited performance, due to the boundary discontinuity issue and the ambiguities in sub-crack orientation. In this paper, we propose a first-of-its-kind oriented sub-crack detector, dubbed as CrackDet, which is derived from a novel piecewise angle definition, to ease the boundary discontinuity problem. And then, we propose a multi-branch angle regression loss for learning sub-crack orientation and variance together. Since there are no related benchmarks, we construct three fully annotated datasets, namely, ORC, ONPP, and OCCSD, which involve various cracks in road pavement and industrial facilities. Experiments show that our approach outperforms state-of-the-art crack detectors.

This work aims to develop a framework of automatic medical fraud detection (AMFD) which can be deployed in healthcare industry. To address the issue that the medical fraud labels are insufficient in both size and classes for training a good AMFD model, this work proposes a novel Variational AutoEncoder-based Relational Model (VAERM) which can simultaneously exploit Patient-Doctor relational network and one-class fraud labels to improve the fraud detection. Then, the proposed VAERM coupled with active learning strategy can assist healthcare industry experts to conduct cost-effective fraud investigation. Finally, we propose an online model updating method to reduce the computation and memory requirement while preserving the predictive performance. The proposed framework is tested in a real world dataset and it empirically outperforms the state-of-the-art methods in both automatic fraud detection and fraud investigation tasks.

Ubiquitylation is a critical post-translational modification (PTM) process that performs a critical role in a wide range of biological functions and is closely related to hallmarks of cancer, such as DNA damage response and oxidative stress. Over the past several years, deep learning have been widely employed in protein ubiquitylation site prediction tools. However, existing deep learning tools have a common feature that they treat protein sequences as input without considering spatial information of protein. This work exploits the three-dimensional structural protein to develop a novel graph-driven ubiquitylation site predictive model (GraphUbiquSite) combing capsule module to improve predictive accuracy. According to the experimental results on Protein Lysine Modification Database (PLMD), the proposed model can archive better performance in comparison with the state-of-the-art methods.

Despite the substantial progress of deep models for crack recognition, due to the inconsistent cracks in varying sizes, shapes, and noisy background textures, there still lacks the discriminative power of the deeply learned features when supervised by the cross-entropy loss. In this paper, we propose the geometry-aware guided loss (GAGL) that enhances the discrimination ability and is only applied in the training stage without extra computation and memory during inference. The GAGL consists of the feature-based geometry-aware projected gradient descent method (FGA-PGD) that approximates the geometric distances of the features to the class boundaries, and the geometry-aware update rule that learns an anchor of each class as the approximation of the feature expected to have the largest geometric distance to the corresponding class boundary. Then the discriminative power can be enhanced by minimizing the distances between the features and their corresponding class anchors in the feature space. To address the limited availability of related benchmarks, we collect a fully annotated dataset, namely, NPP2021, which involves inconsistent cracks and noisy backgrounds in real-world nuclear power plants. Our proposed GAGL outperforms the state of the arts on various benchmark datasets including CRACK2019, SDNET2018, and our NPP2021.

Recent studies show that introducing phonetic information into multi-task learning could significantly improve the performance of speaker embedding extraction. However, benefits of such architectures usually depend largely on the availibility of a well-matched dataset, and domain or language mismatch would result in obvious dropdown in performance. Meanwhile, the utilization of these massive mismatched data and application of these auxiliary tasks may bring many rich features that could be exploited. In this paper, we propose a phoneme-aware adaptation network with discrepancy minimization and dynamically-classified vector for text-independent speaker verification to address these abovementioned challenges. More specifically, our method first utilize the maximum mean discrepancy (MMD) as part of the total loss function to solve the mismatch between training data of the speaker subnet and the phoneme subnet. And then we use a dynamically-classified vector-guided softmax loss (DV-Softmax), which could adaptively emphasize different high-quality features and dynamically change their weights, to guide the discriminative speaker embedding. Experimental results on VoxCeleb1 data set confirmed its superiority against the other state-of-the-art phoneme adaptation methods, providing approximately 15% relative improvements in equal error rate (EER).

MOTIVATION: Interaction between transcription factor (TF) and its target genes establishes the knowledge foundation for biological researches in transcriptional regulation, the number of which is, however, still limited by biological techniques. Existing computational methods relevant to the prediction of TF-target interactions are mostly proposed for predicting binding sites, rather than directly predicting the interactions. To this end, we propose here a graph attention-based autoencoder model to predict TF-target gene interactions using the information of the known TF-target gene interaction network combined with two sequential and chemical gene characters, considering that the unobserved interactions between transcription factors and target genes can be predicted by learning the pattern of the known ones. To the best of our knowledge, the proposed model is the first attempt to solve this problem by learning patterns from the known TF-target gene interaction network. RESULTS: In this paper, we formulate the prediction task of TF-target gene interactions as a link prediction problem on a complex knowledge graph and propose a deep learning model called GraphTGI, which is composed of a graph attention-based encoder and a bilinear decoder. We evaluated the prediction performance of the proposed method on a real dataset, and the experimental results show that the proposed model yields outstanding performance with an average AUC value of 0.8864 +/- 0.0057 in the 5-fold cross-validation. It is anticipated that the GraphTGI model can effectively and efficiently predict TF-target gene interactions on a large scale. AVAILABILITY: Python code and the datasets used in our studies are made available at https://github.com/YanghanWu/GraphTGI.

MOTIVATION: Identifying the target genes of transcription factors (TFs) is of great significance for biomedical researches. However, using biological experiments to identify TF-target gene interactions is still time consuming, expensive and limited to small scale. Existing computational methods for predicting underlying genes for TF to target is mainly proposed for their binding sites rather than the direct interaction. To bridge this gap, we in this work proposed a deep learning prediction model, named HGETGI, to identify the new TF-target gene interaction. Specifically, the proposed HGETGI model learns the patterns of the known interaction between TF and target gene complemented with their involvement in different human disease mechanisms. It performs prediction based on random walk for meta-path sampling and node embedding in a skip-gram manner. RESULTS: We evaluated the prediction performance of the proposed method on a real dataset and the experimental results show that it can achieve the average area under the curve of 0.8519 ± 0.0731 in fivefold cross validation. Besides, we conducted case studies on the prediction of two important kinds of TF, NFKB1 and TP53. As a result, 33 and 32 in the top-40 ranking lists of NFKB1 and TP53 were successfully confirmed by looking up another public database (hTftarget). It is envisioned that the proposed HGETGI method is feasible and effective for predicting TF-target gene interactions on a large scale. AVAILABILITY AND IMPLEMENTATION: The source code and dataset are available at https://github.com/PGTSING/HGETGI. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Unmanned aerial vehicles (UAVs) are being used in different emerging domains for accomplishing many critical tasks. However, due to the various constraints, such as battery life, computational resources, etc., a UAV under a mission (M-UAV) often needs assistance from an edge/cloud server that is reachable from the M-UAV’s location. A connection between an M-UAV and edge server can be established via an access point or AP. Therefore, before sharing any sensitive information with the edge server, it is essential for an M-UAV to determine the legitimacy of the selected AP. Recently, some works in this direction indicate that a rogue UAV (R-UAV) can successfully mimic a legitimate AP for intercepting the communication channel. Hence, there should be a robust detection mechanism in place for addressing such a threat scenario. In this article, considering one of the emerging domains—the Internet of Vehicle (IoV) networks, at first, we show that communication in the IoV networks can get benefit from the presence of M-UAVs. However, as the link between the M-UAV and edge server can be intercepted by an R-UAV, the adversary may access the sensitive information from the IoV networks. Followed by this, we propose atiming-basedalgorithm for identifying the presence of rogue APs (or R-UAVs) in the channel. The M-UAV executes the timing-based algorithm, and the detection methoddoes notrequire any auxiliary hardware or any modification to the network protocols for meeting the objective. Supported by an extensive evaluation study, we show that without any rigid restriction on the M-UAV’s speed (e.g., by limiting it to almost static) the proposed approach significantly enhances the detection accuracy (at least by a margin of 29.7% and 16.65%) compared to the state-of-the-art methods.

Colonoscopy screening is one of the most effective diagnostic tools for detecting intestinal diseases, such as bleeding, polyp, Meckel’s diverticulum, and ulcer. However, the missed rate of manual detection is high due to the lack of experience or fatigue among clinicians. To address this issue, this work proposed a novel autodiagnosis framework built on Internet of Medical Things (IoMT) systems, which can be deployed among multiple hospitals in a distributed fusion. This work presents a two-stage knowledge distillation (TSKD) method coupled with Bayesian optimization (BO) that can exploit distributed colonoscopy data to learn a compact diagnostic model achieving a good tradeoff between predictive performance and resource consumption (e.g., memory and computation). The proposed framework is extensively evaluated in real-world data sets in comparison with its counterparts. A prototype of on-site diagnostic device is implemented to demonstrate the potential for real-world deployment.

With unmanned aerial vehicle (UAV) technologies advanced rapidly, many applications have emerged in cities. However, those applications do not widely spread as the safety consideration hinders the UAV from integrating into the civilian environment. This work focuses on investigating the UAV emergency landing problem which is a critical safety functionality of UAV. This work proposed a graph convolution network (GCN)-based decision network to learn by imitating the human pilots’ landing strategy. To alleviate the needs of a large amount of real-world data for model training, the proposed model allows to be trained in a simulated environment and then transferred to the real-world scenario due to the separation of domain-specific terrain classes and domain-independent topological structures among down-looking camera images. The GCN-based decision network can be coupled with a topological heuristic to improve the performance of action prediction in an emergency situation. To evaluate the proposed method, this work implemented a simulation environment for collecting data and testing the UAV emergency landing. The empirical results in both simulated and real-world scenarios show that the proposed methods can outperform the state-of-the-art counterparts in terms of predictive accuracy and success landing rate.

The studies of Unmanned Air/Ground Vehicle (UAV/UGV) cooperative detection systems have received much attention due to their wide applications in the disaster rescue, target tracking, intelligent surveillance, and automatical package delivery missions. UAVs provide a broad view and have a fast speed in the air, while UGVs have sufficient load capacity and can serve as repeater stations on the ground. The path planning of a UAV/UGV cooperative system is an important but difficult issue, which aims to plan paths for both the UAVs and the UGVs in the system to cooperatively complete a mission. In this article, we consider the path planning problem of the UAV/UGV cooperative system for illegal urban building detection, by taking the limits of UGV speed, UAV load power, and UAV/UGV communication restriction into consideration. To solve this problem, we first model the path planning problem as a constraint optimization problem which tries to minimize an overall execution time for completing the illegal urban building detection tasks, and then propose a two-level memetic algorithm (called Two-MA) to solve the path planning problems of both the UAV and the UGV. Experiments on both synthetic and real-world data sets show the superiority of the proposed Two-MA over several states-of-the-art algorithms in solving the path planning problems of the UAV and UGV for illegal urban building detection tasks. Note to Practitioners—This article was motivated by the task of detecting illegal buildings in cities by unmanned vehicles. Previous works mainly focus on path planning of either UAVs or UGVs in this task. This article proposes a new approach using an Unmanned Air/Ground Vehicle (UAV/UGV) cooperative system for detecting illegal buildings in parks, by taking the limits of UGV speed, UAV load power, and UAV/UGV communication restriction into consideration. This cooperative system consists of a UAV, UGV, and control center. The UAV equipped with cameras takes aerial photography in the air, and can transmit collected photos to the control center. The UGV executes loading and transportation on the ground, and can serve as takeoff and landing platforms for the UAV. The control center executes computationally intensive tasks such as data transmission and processing, task scheduling, and vehicle coordination. To quickly complete all detection tasks, a memetic algorithm is proposed for path planning of both the UAV and the UGV. The simulated results show that the proposed algorithm enables the UAV/UGV cooperative system to visit all buildings in cities with a minimum task execution time.

Auscultation is an important tool for diagnosing respiratory-related diseases. Unfortunately, the quality of auscultation is limited by the professional level of the doctor and the environment of the auscultation. Some studies have focused on automated auscultation techniques. However, existing approaches suffer from two challenges: 1) the models cannot learn from data distributed among multiple hospitals and 2) the predictions of the models are difficult to interpret for physicians. To address this issue, this article proposes a novel explainable respiratory sound analysis framework with fuzzy decision tree regularization. This framework develops an ensemble knowledge distillation technique to learn distributed data and achieves good performance in terms of model efficiency and accuracy. Fuzzy decision trees are used to explain the predictions of the model and produce decision rules that can be well accepted by physicians. The effectiveness of this framework is thoroughly validated on the Respiratory Sound database and compared with other existing approaches.

Stroke has become a leading cause of death and long-term disability in the world with no effective treatment. Deep learning-based approaches have the potential to outperform existing stroke risk prediction models, but they rely on large well-labeled data. Due to the strict privacy protection policy in health-care systems, stroke data is usually distributed among different hospitals in small pieces. In addition, the positive and negative instances of such data are extremely imbalanced. Transfer learning can solve small data issue by exploiting the knowledge of a correlated domain, especially when multiple source of data are available. In this work, we propose a novel Hybrid Deep Transfer Learning-based Stroke Risk Prediction (HDTL-SRP) scheme to exploit the knowledge structure from multiple correlated sources (i.e., external stroke data, chronic diseases data, such as hypertension and diabetes). The proposed framework has been extensively tested in synthetic and real-world scenarios, and it outperforms the state-of-the-art stroke risk prediction models. It also shows the potential of real-world deployment among multiple hospitals aided with 5 G/B5G infrastructures.

Visual classification algorithms based-on Deep Neural Networks (DNN) have been widely adopted in autonomous vehicle design. However, DNN suffers from adversarial attacks including pixel attacks and patch attacks, and its adoption may introduce new vulnerability into such security-critical scenarios. Existing defense techniques only focus on defending against one category, either pixel attacks or patch attacks, but does not translate to the other. Hence, the design of a practical comprehensive real-time defense algorithm for DNN-based classifiers presents a challenging task in this adversarial context. This paper attempts to address the abovementioned problem by combining Compressive Sensing with Generative neural networks (CSG) to construct an efficient defense framework, in conjunction with the proposal of a classifier-aware adversarial training way. Extensive experiments have been conducted using the LISA road sign dataset to evaluate the performance of CSG. The results show its superiority in comprehensively defending adversarial examples generated using attacks including CW-L2, FGSM and Sticker, compared with other state-of-the-art defense techniques.

In a collaborative air-ground robotic system, the large-scale terrain mapping using aerial images is important for the ground robot to plan a globally optimal path. However, it is a challenging task in a novel and dynamic field without historical human supervision. To alleviate the reliance on human intervention, this article presents a novel framework that integrates active learning and generative adversarial networks (GANs) to effectively exploit small human-labeled data for terrain mapping. In order to model the diverse terrain patterns, this article designs two novel diversity-sensitive GAN models which can capture fine-grained terrain classes among aerial image patches. The proposed approaches are tested in two real-world scenarios using our collaborative air-ground robotic platform. The empirical results show that our methods can outperform their counterparts in the predictive accuracy of terrain classification, visual quality of terrain mapping, and average length of the planned ground path. In practice, the proposed terrain mapping framework is especially valuable when the budget in time or labor cost is very limited.

This paper presents a variational Bayesian kernel selection (VBKS) algorithm for sparse Gaussian process regression (SGPR) models. In contrast to existing GP kernel selection algorithms that aim to select only one kernel with the highest model evidence, our VBKS algorithm considers the kernel as a random variable and learns its belief from data such that the uncertainty of the kernel can be interpreted and exploited to avoid overconfident GP predictions. To achieve this, we represent the probabilistic kernel as an additional variational variable in a variational inference (VI) framework for SGPR models where its posterior belief is learned together with that of the other variational variables (i.e., inducing variables and kernel hyperparameters). In particular, we transform the discrete kernel belief into a continuous parametric distribution via reparameterization in order to apply VI. Though it is computationally challenging to jointly optimize a large number of hyperparameters due to many kernels being evaluated simultaneously by our VBKS algorithm, we show that the variational lower bound of the log-marginal likelihood can be decomposed into an additive form such that each additive term depends only on a disjoint subset of the variational variables and can thus be optimized independently. Stochastic optimization is then used to maximize the variational lower bound by iteratively improving the variational approximation of the exact posterior belief via stochastic gradient ascent, which incurs constant time per iteration and hence scales to big data. We empirically evaluate the performance of our VBKS algorithm on synthetic and massive real-world datasets.

The radio frequency identification (RFID) technologies, which have been widely adopted in different Internet of Things (IoT) applications, are the fundamental building block for achieving smart factories, smart logistics, smart stores, etc. Besides knowing the ID of tags, the relative location information of different tags is of great importance since it contains the spatial relationship among different tags, which is essential for many object localization applications beyond the capacity of absolute localization approaches. In this article, we proposeTagSort, an RFID-based sorting system that exploits the physical layer information, i.e., the phase of RFID wireless signals to achieve the relative localization of different tags. Several novel filtering and peak detection algorithms are proposed to achieve accurate and robust detection of the order of tags. Extensive evaluation shows promising results (over 95% accuracy) and makeTagSorta promising system for future RFID sorting systems, thus enabling a variety of IoT applications and services.

The expressive power of a Gaussian process (GP) model comes at a cost of poor scalability in the data size. To improve its scalability, this paper presents a low-rank-cum-Markov approximation (LMA) of the GP model that is novel in leveraging the dual computational advantages stemming from complementing a low-rank approximate representation of the full-rank GP based on a support set of inputs with a Markov approximation of the resulting residual process; the latter approximation is guaranteed to be closest in the Kullback-Leibler distance criterion subject to some constraint and is considerably more refined than that of existing sparse GP models utilizing low-rank representations due to its more relaxed conditional independence assumption (especially with larger data). As a result, our LMA method can trade off between the size of the support set and the order of the Markov property to (a) incur lower computational cost than such sparse GP models while achieving predictive performance comparable to them and (b) accurately represent features/patterns of any scale. Interestingly, varying the Markov order produces a spectrum of LMAs with PIC approximation and full-rank GP at the two extremes. An advantage of our LMA method is that it is amenable to parallelization on multiple machines/cores, thereby gaining greater scalability. Empirical evaluation on three real-world datasets in clusters of up to 32 computing nodes shows that our centralized and parallel LMA methods are significantly more time-efficient and scalable than state-of-the-art sparse and full-rank GP regression methods while achieving comparable predictive performances.

Mobility-on-demand (MoD) systems have recently emerged as a promising paradigm of one-way vehicle sharing for sustainable personal urban mobility in densely populated cities. We assume the capability of a MoD system to be enhanced by deploying robotic shared vehicles that can autonomously cruise the streets to be hailed by users. A key challenge of the MoD system is that of real-time, fine-grained mobility demand and traffic flow sensing and prediction. This paper presents novel Gaussian process (GP) decentralized data fusion and active sensing algorithms for real-time, fine-grained traffic modeling and prediction with a fleet of MoD vehicles. The predictive performance of our decentralized data fusion algorithms are theoretically guaranteed to be equivalent to that of sophisticated centralized sparse GP approximations. We derive consensus filtering variants requiring only local communication between neighboring vehicles. We theoretically guarantee the performance of our decentralized active sensing algorithms. When they are used to gather informative data for mobility demand prediction, they can achieve a dual effect of fleet rebalancing to service mobility demands. Empirical evaluation on real-world datasets shows that our algorithms are significantly more time-efficient and scalable in the size of data and fleet while achieving predictive performance comparable to that of state-of-the-art algorithms. Note to Practitioners-Knowing, understanding, and predicting spatiotemporally varying traffic phenomena in real time has become increasingly important to the goal of achieving smooth-flowing, congestion-free traffic in densely populated urban cities, which motivates our work here. This paper addresses the following fundamental problem of data fusion and active sensing: How can a fleet of autonomous robotic vehicles or mobile probes actively cruise a road network to gather and assimilate the most informative data for predicting a spatiotemporally varying traffic phenomenon like a mobility demand pattern or traffic flow? Existing centralized solutions are poorly suited because they suffer from a single point of failure and incur huge communication, space, and time overheads with large data and fleet. This paper proposes novel efficient and scalable decentralized data fusion and active sensing algorithms with theoretical performance guarantees. The practical applicability of our algorithms is not restricted to traffic monitoring [1]-[4]; they can be used in other environmental sensing applications such as mineral prospecting [5], precision agriculture, monitoring of ocean/freshwater phenomena (e.g., plankton bloom) [6]-[9], forest ecosystems, pollution (e.g., oil spill), or contamination. Note that the decentralized data fusion component of our algorithms can also be used for static sensors and passive mobile probes and, interestingly, adapted to parallel implementations to be run on a cluster of machines for achieving efficient and scalable probabilistic prediction (i.e., with predictive uncertainty) with large data. Empirical results show that our algorithms can perform well with two datasets featuring real-world traffic phenomena in the densely-populated urban city of Singapore. A limitation of our algorithms is that the decentralized data fusion components assume independence between multiple traffic phenomena while the decentralized active sensing components only work for a single traffic phenomenon. So, in our future work, we will generalize our algorithms to perform active sensing of multiple traffic phenomena and remove the assumption of independence between them.

Central to robot exploration and mapping is the task of persistent localization in environmental fields characterized by spatially correlated measurements. This paper presents a Gaussian process localization (GP-Localize) algorithm that, in contrast to existing works, can exploit the spatially correlated field measurements taken during a robot's exploration (instead of relying on prior training data) for efficiently and scalably learning the GP observation model online through our proposed novel online sparse GP. As a result, GP-Localize is capable of achieving constant time and memory (i.e., independent of the size of the data) per filtering step, which demonstrates the practical feasibility of using GPs for persistent robot localization and autonomy. Empirical evaluation via simulated experiments with real-world datasets and a real robot experiment shows that GP-Localize outperforms existing GP localization algorithms.