Zeroth-order (ZO) optimization has gained attention as a memory-efficient alternative to first-order (FO) methods, particularly in settings where gradient computation is expensive or even impractical. Beyond its memory efficiency, in this work, we investigate ZO optimization for continual learning (CL) as a novel approach to address the plasticity-stability-efficiency trilemma. Through theoretical analysis and empirical evidence, we show that ZO optimization naturally leads to flatter loss landscapes, which in turn reduce forgetting in CL. However, this stability comes at a cost of plasticity: due to its imprecise gradient estimates and slower convergence, ZO optimization tends to be less effective than FO in acquiring new task-specific knowledge, particularly under constrained training budgets. To better understand this trade-off, we conduct a holistic evaluation of ZO optimization applied to various existing CL methods. Our findings reveal that ZO optimization enhances stability but often undermines plasticity, particularly when used with learnable classifiers. Motivated by this insight, we propose ZO-FC, a simple but effective approach that applies ZO optimization to a single adapter-based PEFT module with FO optimized classifier. This design leverages the stability benefits of ZO while preserving the adaptability of FO updates with negligible memory overhead. Experiments demonstrate that ZO-FC achieves an effective balance between stability and plasticity, offering a practical and memory-efficient solution for on-device CL.

Temporal Graph Neural Networks (TGNNs) have become indispensable for analyzing dynamic graphs in critical applications such as social networks, communication systems, and financial networks. However, the robustness of TGNNs against adversarial attacks, particularly sophisticated attacks that exploit the temporal dimension, remains a significant challenge. Existing attack methods for Spatio-Temporal Dynamic Graphs (STDGs) often rely on simplistic, easily detectable perturbations (e.g., random edge additions/deletions) and fail to strategically target the most influential nodes and edges for maximum impact. We introduce the High Impact Attack (HIA), a novel restricted black-box attack framework specifically designed to overcome these limitations and expose critical vulnerabilities in TGNNs. HIA leverages a data-driven surrogate model to identify structurally important nodes (central to network connectivity) and dynamically important nodes (critical for the graph's temporal evolution). It then employs a hybrid perturbation strategy, combining strategic edge injection (to create misleading connections) and targeted edge deletion (to disrupt essential pathways), maximizing TGNN performance degradation. Importantly, HIA minimizes the number of perturbations to enhance stealth, making it more challenging to detect. Comprehensive experiments on five real-world datasets and four representative TGNN architectures (TGN, JODIE, DySAT, and TGAT) demonstrate that HIA significantly reduces TGNN accuracy on the link prediction task, achieving up to a 35.55% decrease in Mean Reciprocal Rank (MRR) - a substantial improvement over state-of-the-art baselines. These results highlight fundamental vulnerabilities in current STDG models and underscore the urgent need for robust defenses that account for both structural and temporal dynamics. Code and Data are available at https://github.com/ryandhjeon/hia.

In this study, we developed an automatic pipeline for molecular dynamics (MD) simulations and a deep learning model for predicting protein-protein binding affinities. The framework will output common post-simulation analyses, including total energy, Root Mean Square Deviation (RMSD), Root Mean Square Fluctuation (RMSF), hydrogen bonds, and salt bridges. It will minimize manual interaction. This tool also integrates a 3D Convolutional Neural Network (CNN) that processes structural features from simulation trajectories to predict binding affinity (pKd and$\Delta G$). We evaluate several model versions with different feature map sizes to balance between predictive accuracy wit hardware requirements. A case study on the human urate transporter GLUT9 (PDB: 8Y65) demonstrates the pipeline's ability to provide rapid, initial insights into protein stability and dynamics. This automated approach serves as a powerful complement to traditional analysis tools. It allows researchers to quickly screen simulation results and identify systems for more detailed investigation.

While deep-learning-enabled recommender systems demonstrate strong performance benchmarks, many struggle to adapt effectively in real-world environments due to limited use of user-item relationship data and insufficient transparency in recommendation generation. Traditional collaborative filtering approaches fail to integrate multifaceted item attributes, and although Factorization Machines account for item-specific details, they overlook broader relational patterns. Collaborative knowledge graph-based models have progressed by embedding user-item interactions with item-attribute relationships, offering a holistic perspective on interconnected entities. However, these models frequently aggregate attribute and interaction data in an implicit manner, leaving valuable relational nuances underutilized.This study introduces the Knowledge Graph Attention Network with Information Fusion (KGIF), a specialized framework designed to merge entity and relation embeddings explicitly through a tailored self-attention mechanism. The KGIF framework integrates reparameterization via dynamic projection vectors, enabling embeddings to adaptively represent intricate relationships within knowledge graphs. This explicit fusion enhances the interplay between user-item interactions and item-attribute relationships, providing a nuanced balance between user-centric and item-centric representations. An attentive propagation mechanism further optimizes knowledge graph embeddings, capturing multi-layered interaction patterns. The contributions of this work include an innovative method for explicit information fusion, improved robustness for sparse knowledge graphs, and the ability to generate explainable recommendations through interpretable path visualization. The implementation and datasets for this study are publicly available1.

As a special type of Knowledge Graph (KG), Continual Knowledge Graph Learning (CKGL) plays a pivotal role in various areas such as recommendation systems, search engines, and personalized services, where knowledge dynamically evolves. A challenging problem in CKGL is catastrophic forgetting, where models forget previously learned knowledge upon being trained on new data. To overcome the challenge, this study proposes Flexible Memory Rotation (FMR), a dual-level regularization technique that focuses on both parameter level and structural level. Our idea is inspired by the natural human learning process, which tends to memorize correctly learned knowledge, leverage the learned to acquire new knowledge, and refine incorrectly learned knowledge with newly obtained information. Commonly, existing regularization-based methods fail to mimic this human nature by having a fixed constraint strategy for all model parameters. To this end, the proposed FMR offers flexible constraints based on qualities of learned knowledge evaluated by the Fisher Information Matrix (FIM). Additionally, we identified a limitation of FIM in CKGL, which is the assumption of independence of time steps does not always hold. To overcome this, FMR rotates the parameter space to diagonalize the FIM. This work has four major contributions: 1) develop a novel regularization technique, FMR, a flexible regularization technique, 2) reveal the unexpected failure of FIM in CKGL and provides an easy remedy via parameter space rotation, 3) the comparison experiments on four benchmark datasets designed for CKGL demonstrates improvement over various state-of-the-art (SOTA) CKGL models, and 4) a comprehensive ablation study investigates each component of the proposed model. The source code is available at https://github.com/lijingzhu1/FMR.

Binding affinity (BA) prediction is important for drug discovery and protein engineering. It seeks to understand the interaction strength between proteins and their ligands (or proteins). This information assists in the design of proteins with enhanced or novel functions, as well as understanding the molecular mechanisms of drug action. This paper presents the development and comparative analysis of two deep learning models, a convolutional neural network (CNN) and a transformer model. Many variants of models in this research were developed using TensorFlow. One model that utilizes ProteinBERT was developed using PyTorch. The CNN model captures local sequence features effectively, while the Transformer model leverages self-attention mechanisms to learn long-range dependencies within the sequences. Protein sequences are the inputs for the models. The sequences are processed using various encoders, like One-hot encoding, Sequence-Statistics-Content, and Position Specific Scoring Matrix. The predicted outputs are Gibbs free energy changes, a key indicator of binding affinity. From this study, both the CNN and transformer models can achieve the same level of accuracy under different conditions. For the CNN model, it can handle full data without sacrificing performance, but it takes much more time to preprocess the features from the protein sequences. The transformer model can achieve the same level of accuracy as the CNN model with no big predictive errors for each protein, but it requires the model to run on less data, which removes some rarely long protein sequences. This study emphasizes the potential of advanced deep learning architectures to enhance the predictive strengths of binding affinity models.

Deep Learning (DL) is being applied in various domains, especially in safety-critical applications such as autonomous driving. Consequently, it is of great significance to ensure the robustness of these methods and thus counteract uncertain behaviors caused by adversarial attacks. In this paper, we use gradient heatmaps to analyze the response characteristics of the VGG-16 model when the input images are mixed with adversarial noise and statistically similar Gaussian random noise. In particular, we compare the network response layer by layer to determine where errors occurred. Several interesting findings are derived. First, compared to Gaussian random noise, intentionally generated adversarial noise causes severe behavior deviation by distracting the area of concentration in the networks. Second, in many cases, adversarial examples only need to compromise a few intermediate blocks to mislead the final decision. Third, our experiments revealed that specific blocks are more vulnerable and easier to exploit by adversarial examples. Finally, we demonstrate that the layers Block4_conv1 and Block5_ cov1 of the VGG-16 model are more susceptible to adversarial attacks. Our work could potentially provide useful insights into developing more reliable Deep Neural Network (DNN) models.

Despite the recent state-of-the-art performance of Deep Learning (DL), imbalanced graph-structured data remains an open challenge in social science, traffic networks, and biomedical informatics. Recently, a surge in research on Representation Learning has significantly improved the performance of DL algorithms on imbalanced non-graph-structured data. In addition, Graph Neural Networks (GNNs) already in widespread use for representing graph-structured data in DL models with more advanced techniques in neural message-passing and deep graph embedding. However, most existing works are based on assumptions that oversimplify the complexity of real-world problems. In this paper, we propose Reinforced Contrastive GNNs (RCGNN), a novel graph representation learning model for anomaly detection with multi-relational graph-structured data. The proposed model produces a neighbor selection with Reinforcement Learning (RL) based on the similarity of neighborhoods in multi-relational structured graphs. In addition, the graph representation is learned by an adaptive AutoEncoder (AE) with Triplet Loss (TL) in Contrastive Learning. By aggregating the nodes with the highest similarities in their features and the importance of each node, our model is able to construct the multi-relational graphs by keeping the complexity of the graph structure as well as the relation-dependency representations. Experiments on multiple benchmark data sets demonstrate the advantage of RCGNN in learning better representations for multi-relational graphs. Furthermore, compared to other GNN models, our model shows better performance in accuracy, F1, and PR AUC scores.

Big data analytics and mining have the potential to enable real-time decision making and control in a range of Internet of Things (IoT) application domains, such as the Internet of Vehicles, the Internet of Wings, and the Airport of Things. The prediction toward air mobility, which is essential to the studies of air traffic management, has been a challenging task due to the complex spatial and temporal dependencies in air traffic data with highly nonlinear and variational patterns. Existing works for air traffic prediction only focus on either modeling static traffic patterns of individual flight or temporal correlation, with no or limited addressing of the spatial impact, namely, the propagation of traffic perturbation among airports. In this article, we propose to leverage the concept of graph and model the airports as nodes with time-series features and conduct data mining on graph-structured data. To be specific, first, airline on-time performance (AOTP) data is preprocessed to generate a temporal graph data set, which includes three features: 1) the number; 2) average delay; and 3) average taxiing time of departure and arrival flights. Then, a spatial–temporal graph neural networks model is implemented to forecast the mobility level at each airport over time, where a combination of graph convolution and time-dimensional convolution is used to capture the spatial and temporal correlation simultaneously. Experiments on the data set demonstrate the advantage of the model on spatial–temporal air mobility prediction, together with the impact of different priors on adjacency matrices and the effectiveness of the temporal attention mechanism. Finally, we analyze the prediction performance and discuss the capability of our model. The prediction framework proposed in this work has the potential to be generalized to other spatial–temporal tasks in IoT.

The Internet of Things (IoT) is becoming an indispensable part of everyday life, enabling a variety of emerging services and applications. However, the presence of rogue IoT devices has exposed the IoT to untold risks with severe consequences. The first step in securing the IoT is detecting rogue IoT devices and identifying legitimate ones. Conventional approaches use cryptographic mechanisms to authenticate and verify legitimate devices’ identities. However, cryptographic protocols are not available in many systems. Meanwhile, these methods are less effective when legitimate devices can be exploited or encryption keys are disclosed. Therefore, noncryptographic IoT-device identification and rogue device detection become efficient solutions to secure existing systems and will provide additional protection to systems with cryptographic protocols. Noncryptographic approaches require more effort and are not yet adequately investigated. In this article, we provide a comprehensive survey on machine learning technologies for the identification of IoT devices along with the detection of compromised or falsified ones from the viewpoint of passive surveillance agents or network operators. We classify the IoT-device identification and detection into four categories: 1) device-specific pattern recognition; 2) deep learning-enabled device identification; 3) unsupervised device identification; and 4) abnormal device detection. Meanwhile, we discuss various ML-related enabling technologies for this purpose. These enabling technologies include learning algorithms, feature engineering on network traffic traces and wireless signals, incremental learning, and abnormality detection.

Abnormal event detection with the lowest latency is an indispensable function for safety-critical systems, such as cyber defense systems. However, as systems become increasingly complicated, conventional sequential event detection methods become less effective, especially when we need to define indicator metrics from complicated data manually. Although deep neural networks (DNNs) have been used to handle heterogeneous data, the theoretic assurability and explainability are still insufficient. This article provides a holistic framework for the quickest and sequential detection of abnormalities and time-dependent abnormal events. We explore the latent space characteristics of zero-bias neural networks considering the classification boundaries and abnormalities. We then provide a novel method to convert zero-bias DNN classifiers into performance-assured binary abnormality detectors. Finally, we provide a sequential quickest detection (QD) scheme that provides the theoretically assured lowest abnormal event detection delay under false alarm constraints using the converted abnormality detector. We verify the effectiveness of the framework using real massive signal records in aviation communication systems and simulation. Codes and data are available athttps://github.com/pcwhy/AbnormalityDetectionInZbDNN.

The ubiquitous deployment of 5G New Radio (5G NR) stimulates Unmanned Aircraft Systems (UAS) swarm networking to evolve to achieve more imminent progress. The heterogeneous collaboration between UAS swarm enhances the complexity and the efficiency of mission complement that requires robustness, flexibility, and sustainability of throughput in UAS swarm networking. The conventional approaches mainly are based on the hierarchical architectures that are limited to satisfy the challenges of UAS swarm with high dynamics on a large scale. In this paper, we propose an optimal cell wall paradigm to enhance the throughput in heterogeneous UAS swarm networking. With the weight adjustment of each link, we map the optimization into a polyhedron scheduling problem and formula the problem into Max-min Throughput Fair Scheduling (MTFS). Further, we propose a max-min throughput algorithm to optimize the minimum throughput of cell wall paradigm. With the optimal max-min throughput, we optimize the schedule with edge-coloring to achieve global MTFS solving. The normalized MTFS shows our algorithm can achieve over 40% improvement of MTFS globally. In terms of MTFS solving, our algorithms have promising potential to improve the throughput and mitigate the incidents for multiple beams enabling of UAS in cell wall communication. With the throughput enhancement, the advanced aerial mobility of UAS swarm networking can be escalated on a large scale.

The ubiquitous of 5G New Radio (5G NR) accelerates the massive implementations in many fields including swarm Unmanned Aircraft System (UAS) networking. The ultra capacities of 5G NR can provide more sufficient networking services for the swarm UAS networking which can enable swarm UAS to deploy in more complex and challenging scenarios to achieve missions. However, the conventional swarm UAS networking are mainly centralized or hierarchical which is vulnerable to the dynamics and the deployment of swarm UAS networking on a large scale. In this paper, we formulate a cell wall communications for the heterogeneous swarm UAS networking with the inspiration of biological cell wall communication. Fueled by reinforcement learning, we resolve the edge-coloring problem of cell wall communication scheduling to achieve the maximum throughput between the heterogeneous swarm UAS networking globally. The evaluation shows our proposed reinforcement learning enabled algorithm can surpass the conventional scheduling algorithms over 90% when the time piece is less than 0.01s and achieve the optimal throughput for the heterogeneous swarm UAS networking.

The Internet of Things (IoT) is reshaping modern society by allowing a decent number of RF devices to connect and share information through RF channels. However, such an open nature also brings obstacles to surveillance. For alleviation, a surveillance oracle, or a cognitive communication entity needs to identify and confirm the appearance of known or unknown signal sources in real-time. In this paper, we provide a deep learning framework for RF signal surveillance. Specifically, we jointly integrate the Deep Neural Networks (DNNs) and Quickest Detection (QD) to form a sequential signal surveillance scheme. We first analyze the latent space characteristic of neural network classification models, and then we leverage the response characteristics of DNN classifiers and propose a novel method to transform existing DNN classifiers into performance-assured binary abnormality detectors. In this way, we seamless integrate the DNNs with parametric quickest detection. Finally, we propose an enhanced Elastic Weight Consolidation (EWC) algorithm with better numerical stability for DNNs in signal surveillance system to evolve incrementally, we demonstrate that the zero-bias DNN is superior than regular DNN models considering incremental learning and decision fairness. We evaluated the proposed framework using real signal datasets and we believe this framework is helpful in developing a trustworthy IoT ecosystem.

With the ubiquitous deployment of 5G cellular networking in many fields, unmanned aerial vehicle (UAV) networking, as one of the main parts of the Internet of Things (IoT), is playing a pivot role in the extension of smart cities. Different from the conventional approaches, the 5G enabled UAV networking can be more capable of multiple and complex mission executions with high requirements of collaborations and incorporation. In this paper, we leverage reinforcement learning based scheduling to optimize the throughput of heterogeneous UAV networking. To improve the throughput of the heterogeneous UAV networking, we focus on the balance for the inter-and intra-networking with the reduction of collisions occurring in the time slots. With reinforcement learning enabled scheduling, we can achieve the optimum selections on link activation and time allocation. Compared with the edge coloring of Karloff, our approach can achieve a higher enhancement on the throughput. The experimental results show that our approach reaches the global optimization when tsand tgare less than 0.01. Generally, DQN achieves 57.58% improvement on average which exceeds Karloff. The proposed approach can improve the throughput of heterogeneous UAV networking significantly.

Deep learning (DL) has been utilized pervasively in the Internet of Things (IoT). One typical application of DL in IoT is device identification from wireless signals, namely, noncryptographic device identification (NDI). However, learning components in NDI systems have to evolve to adapt to operational variations, such a paradigm is termed as incremental learning (IL). Various IL algorithms have been proposed and many of them require dedicated space to store the increasing amount of historical data, and therefore, they are not suitable for IoT or mobile applications. Besides, conventional IL schemes can not provide satisfying performance when historical data are not available. In this article, we address the IL problem in NDI from a new perspective, first, we provide a new metric to measure the degree of topological maturity of DNN models from the degree of conflict of class-specific fingerprints. We discover that an important cause for performance degradation in IL-enabled NDI is owing to the conflict of devices’ fingerprints. Second, we also show that the conventional IL schemes can lead to low topological maturity of DNN models in NDI systems. Thirdly, we propose a new channel separation-enabled IL (CSIL) scheme without using historical data, in which our strategy can automatically separate devices’ fingerprints in different learning stages and avoid potential conflict. Finally, We evaluated the effectiveness of the proposed framework using real data from automatic-dependent surveillance-broadcast (ADS-B), an application of IoT in aviation. The proposed framework has the potential to be applied to accurate identification of IoT devices in a variety of IoT applications and services. Data and code available at IEEE Dataport (DOI: 10.21227/1bxc-ke87) andhttps://github.com/pcwhy/CSIL.

The Internet of Things (IoT) provides applications and services that would otherwise not be possible. However, the open nature of IoT makes it vulnerable to cybersecurity threats. Especially, identity spoofing attacks, where an adversary passively listens to the existing radio communications and then mimic the identity of legitimate devices to conduct malicious activities. Existing solutions employ cryptographic signatures to verify the trustworthiness of received information. In prevalent IoT, secret keys for cryptography can potentially be disclosed and disable the verification mechanism. Noncryptographic device verification is needed to ensure trustworthy IoT. In this article, we propose an enhanced deep learning framework for IoT device identification using physical-layer signals. Specifically, we enable our framework to report unseen IoT devices and introduce the zero-bias layer to deep neural networks to increase robustness and interpretability. We have evaluated the effectiveness of the proposed framework using real data from automatic dependent surveillance-broadcast (ADS-B), an application of IoT in aviation. The proposed framework has the potential to be applied to the accurate identification of IoT devices in a variety of IoT applications and services.

Medical image processing is one of the most important topics in the Internet of Medical Things (IoMT). Recently, deep learning methods have carried out state-of-the-art performances on medical imaging tasks. In this paper, we propose a novel transfer learning framework for medical image classification. Moreover, we apply our method COVID-19 diagnosis with lung Computed Tomography (CT) images. However, well-labeled training data sets cannot be easily accessed due to the disease's novelty and privacy policies. The proposed method has two components: reduced-size Unet Segmentation model and Distant Feature Fusion (DFF) classification model. This study is related to a not well-investigated but important transfer learning problem, termed Distant Domain Transfer Learning (DDTL). In this study, we develop a DDTL model for COVID-19 diagnosis using unlabeled Office-31, Caltech-256, and chest X-ray image data sets as the source data, and a small set of labeled COVID-19 lung CT as the target data. The main contributions of this study are: 1) the proposed method benefits from unlabeled data in distant domains which can be easily accessed, 2) it can effectively handle the distribution shift between the training data and the testing data, 3) it has achieved 96% classification accuracy, which is 13% higher classification accuracy than "non-transfer" algorithms, and 8% higher than existing transfer and distant transfer algorithms.

In this paper, we study a not well-investigated but important transfer learning problem termed Distant Domain Transfer Learning (DDTL). This topic is closely related to negative transfer. Unlike conventional transfer learning problems which assume that the source domain and the target domain are more or less similar to each other, DDTL aims to make efficient transfers even when the domains or the tasks are completely different. As an extreme example in image classification, there are only a sufficient amount of unlabeled images of watches, airplanes, and horses in the source domain, and the target domain only has a small set of labeled human face images. Previously, a few instance-based distant domain transfer algorithms were proposed to deal with this type of binary distant domain image classification problems. Yet most existing algorithms are very task-specific and they are only good at binary classification tasks. In this study, we propose a novel feature-based distant domain transfer learning algorithm, which requires only a tiny set of labeled target data and unlabeled source data from completely different domains. Instead of selecting intermediate instances, we introduced Distant Feature Fusion (DFF), a novel feature selection method, to discover general features cross distant domains and tasks by using convolutional autoencoder with a domain distance measurement as a feature extractor. As the novelty of this study, it can effectively handle both distant domain mutil-class image classification and binary image classification problems. More importantly, it has achieved up to 19% higher classification accuracy than "non-transfer" algorithms, and up to 9% higher than existing distant transfer algorithms.