Xiaokang Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Xiaokang Wang has a total of 164 co-authors. Here are the top co-authors by number of shared publications:
- Laurence Tianruo Yang
School of Computer Science and Artificial Intelligence, Zhengzhou University, Zhengzhou, China
23 shared publications - Lei Ren
Beihang University
19 shared publications - Mohammed Jamal Deen
McMaster University, Canada
17 shared publications - Yuanjun Laili
Beihang University
5 shared publications - Bo Liu
School of Computer Science and Artificial Intelligence, Zhengzhou University, Zhengzhou, China
3 shared publications - Jiabao Dong
School of Automation Science and Electrical Engineering, Beihang University, Beijing, China
3 shared publications
Latest Publications
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 75 (7) , 2447-2460
Dynamic Incremental Tucker Decomposition for Sparse Tensors on Heterogeneous Platforms
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 75 (7) , 2447-2460
Dynamic Incremental Tucker Decomposition for Sparse Tensors on Heterogeneous Platforms
- 14 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP, 1
Dynamic Tanh Enhanced Dual-Branch Mixer for Long-Term Time Series Forecasting
Long-term time series forecasting (LTSF) is critical to the Internet of Things (IoT), supporting proactive decision-making in applications such as smart grids and industrial monitoring. While normalization layers are commonly used in LTSF models to stabilize training, they often introduce significant computational overhead and global synchronization bottlenecks, particularly in large-scale or resource-constrained IoT systems. This raises a fundamental question: can these traditional normalization layers be replaced with a more efficient, lightweight mechanism in LTSF models? To explore this possibility, we propose DyT-Mixer, a dual-branch architecture enhanced by a Dynamic Tanh (DyT) function. As a purely element-wise operation, DyT serves as a lightweight alternative to Layer Normalization (LN) within Multilayer Perceptrons (MLPs), acting as a robust internal numerical stabilization module. The architecture employs a decomposed dual-branch strategy: a Triple-Mixing Operator (TMO) captures local, long-range, and cross-variable dependencies for seasonal components, while a Multi-scale Trend Extractor (MTE) models multi-granularity dynamics. Extensive experiments on eight real-world benchmarks demonstrate that DyT-Mixer achieves competitive accuracy. Notably, by eliminating global statistical synchronization, DyT-Mixer improves forecasting accuracy and reduces end-to-end latency, offering a scalable and hardware-friendly stabilization solution for high-dimensional sensor networks in resource-constrained IoT environments.
- 14 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP, 1
Dynamic Tanh Enhanced Dual-Branch Mixer for Long-Term Time Series Forecasting
Long-term time series forecasting (LTSF) is critical to the Internet of Things (IoT), supporting proactive decision-making in applications such as smart grids and industrial monitoring. While normalization layers are commonly used in LTSF models to stabilize training, they often introduce significant computational overhead and global synchronization bottlenecks, particularly in large-scale or resource-constrained IoT systems. This raises a fundamental question: can these traditional normalization layers be replaced with a more efficient, lightweight mechanism in LTSF models? To explore this possibility, we propose DyT-Mixer, a dual-branch architecture enhanced by a Dynamic Tanh (DyT) function. As a purely element-wise operation, DyT serves as a lightweight alternative to Layer Normalization (LN) within Multilayer Perceptrons (MLPs), acting as a robust internal numerical stabilization module. The architecture employs a decomposed dual-branch strategy: a Triple-Mixing Operator (TMO) captures local, long-range, and cross-variable dependencies for seasonal components, while a Multi-scale Trend Extractor (MTE) models multi-granularity dynamics. Extensive experiments on eight real-world benchmarks demonstrate that DyT-Mixer achieves competitive accuracy. Notably, by eliminating global statistical synchronization, DyT-Mixer improves forecasting accuracy and reduces end-to-end latency, offering a scalable and hardware-friendly stabilization solution for high-dimensional sensor networks in resource-constrained IoT environments.
- 28 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. PP (15513203) , 1-12
CAAD: A Cross-Modal Autoregressive Diffusion Approach for Anomaly Detection in Complex Industrial Processes
Accurate anomaly detection is essential for the safe and stable operation of complex industrial processes, such as magnesium smelting. Although diffusion models have shown strong performance in industrial anomaly detection, thanks to their powerful distribution modeling, traditional diffusion models often suffer from identity mapping under global optimization, causing local high-frequency anomalous textures to be preserved rather than repaired. Moreover, existing cross-modal mechanisms typically assume reliable auxiliary sensor signals, leading to conditional misguidance when sensors fail. To address these issues, we propose a cross-modal autoregressive diffusion framework (CAAD). By adopting an autoregressive diffusion strategy, global denoising is decomposed into local sequence generation, with mandatory local perception constraints effectively preventing anomalous shortcut learning. In addition, we propose a dynamic-gated cross-modal alignment mechanism that adaptively modulates the injection strength of physical context, automatically suppressing unreliable signals during multimodal inconsistencies. Extensive experiments on magnesium smelting benchmarks demonstrate that CAAD achieves competitive performance in both detection accuracy and robustness.
- 28 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. PP (15513203) , 1-12
CAAD: A Cross-Modal Autoregressive Diffusion Approach for Anomaly Detection in Complex Industrial Processes
Accurate anomaly detection is essential for the safe and stable operation of complex industrial processes, such as magnesium smelting. Although diffusion models have shown strong performance in industrial anomaly detection, thanks to their powerful distribution modeling, traditional diffusion models often suffer from identity mapping under global optimization, causing local high-frequency anomalous textures to be preserved rather than repaired. Moreover, existing cross-modal mechanisms typically assume reliable auxiliary sensor signals, leading to conditional misguidance when sensors fail. To address these issues, we propose a cross-modal autoregressive diffusion framework (CAAD). By adopting an autoregressive diffusion strategy, global denoising is decomposed into local sequence generation, with mandatory local perception constraints effectively preventing anomalous shortcut learning. In addition, we propose a dynamic-gated cross-modal alignment mechanism that adaptively modulates the injection strength of physical context, automatically suppressing unreliable signals during multimodal inconsistencies. Extensive experiments on magnesium smelting benchmarks demonstrate that CAAD achieves competitive performance in both detection accuracy and robustness.
- 21 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP (18) , 1
Delayed Backdoor: Let the Trigger Fly for a While in Backdoor Attack on Internet of Things
Since Large Language Models (LLMs) have gained wide attention as the core of AI agents in IoT for conversation and generation tasks, their security issues have become more prominent, especially regarding backdoor attacks. Traditional backdoor attacks often rely on fixed triggers and static outputs, failing to fully exploit the conversational characteristics and generativity of LLMs, which limits their stealth and attack effectiveness in complex human-agent interactions. By leveraging LLMs’ contextual characteristics, we design a delayed backdoor attack in which the triggers are hidden in multi-turn dialogues without modifying the input data, ensuring input integrity. This delayed attack makes the trigger dissociate from poisoned data to enhance stealth and generalization, posing a significant threat to autonomous systems. Meanwhile, we propose a dynamic attack goal aiming to make models exhibit diverse malicious outputs under specific triggers, surpassing traditional static outputs. Experimental results show that our method significantly outperforms traditional methods, achieving a 20% to 80% performance improvement with a 5% poisoning rate and demonstrating stronger robustness under defense mechanisms, highlighting a critical vulnerability in the security of AI agents for IoT.
- 21 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP (18) , 1
Delayed Backdoor: Let the Trigger Fly for a While in Backdoor Attack on Internet of Things
Since Large Language Models (LLMs) have gained wide attention as the core of AI agents in IoT for conversation and generation tasks, their security issues have become more prominent, especially regarding backdoor attacks. Traditional backdoor attacks often rely on fixed triggers and static outputs, failing to fully exploit the conversational characteristics and generativity of LLMs, which limits their stealth and attack effectiveness in complex human-agent interactions. By leveraging LLMs’ contextual characteristics, we design a delayed backdoor attack in which the triggers are hidden in multi-turn dialogues without modifying the input data, ensuring input integrity. This delayed attack makes the trigger dissociate from poisoned data to enhance stealth and generalization, posing a significant threat to autonomous systems. Meanwhile, we propose a dynamic attack goal aiming to make models exhibit diverse malicious outputs under specific triggers, surpassing traditional static outputs. Experimental results show that our method significantly outperforms traditional methods, achieving a 20% to 80% performance improvement with a 5% poisoning rate and demonstrating stronger robustness under defense mechanisms, highlighting a critical vulnerability in the security of AI agents for IoT.
- Open Access
- 1 April 2026
- Published by Tsinghua University Press in Tsinghua Science and Technology
- Vol. 31 (2) , 1151-1169
Frequency Domain Features Based Improving Gradient Descent Optimization for Cyber-Physical-Social Intelligence
Cyber-Physical-Social Intelligence (CPSI) enhances the interpretability of Gradient Descent (GD) optimizers and improves image recognition models. To address the black-box nature of optimizers, we propose a new approach to map the GD optimizer to the frequency domain, allowing analysis of gradient variations at different frequencies. This approach aids in selecting optimal training and optimization strategies, offering a novel solution to the challenge of optimizer interpretability. Specifically, firstly, based on the transfer function of the gradient descent optimizer, the real-frequency characteristic and imaginary-frequency characteristic functions of the optimizer are derived for the first time, which provide a new perspective for the selection of the optimizer and the interpretability of the tuning parameter. Then, based on these characteristic functions, the learning properties of the optimizer in the frequency domain are analyzed for the first time, which provides an important reference for improving the performance of the optimizer in real application problems. Finally, the effectiveness of the frequency domain modulation properties is verified through convex and non-convex optimization problems. Experimental results show that the proposed theory not only improves the recognition accuracy, convergence speed, and stability, but also extends its scope in practical applications.
- 1 April 2026
- Published by Tsinghua University Press in Tsinghua Science and Technology
- Vol. 31 (2) , 1151-1169
Open AccessFrequency Domain Features Based Improving Gradient Descent Optimization for Cyber-Physical-Social Intelligence
Cyber-Physical-Social Intelligence (CPSI) enhances the interpretability of Gradient Descent (GD) optimizers and improves image recognition models. To address the black-box nature of optimizers, we propose a new approach to map the GD optimizer to the frequency domain, allowing analysis of gradient variations at different frequencies. This approach aids in selecting optimal training and optimization strategies, offering a novel solution to the challenge of optimizer interpretability. Specifically, firstly, based on the transfer function of the gradient descent optimizer, the real-frequency characteristic and imaginary-frequency characteristic functions of the optimizer are derived for the first time, which provide a new perspective for the selection of the optimizer and the interpretability of the tuning parameter. Then, based on these characteristic functions, the learning properties of the optimizer in the frequency domain are analyzed for the first time, which provides an important reference for improving the performance of the optimizer in real application problems. Finally, the effectiveness of the frequency domain modulation properties is verified through convex and non-convex optimization problems. Experimental results show that the proposed theory not only improves the recognition accuracy, convergence speed, and stability, but also extends its scope in practical applications.
- 19 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. PP (00983063) , 1
One Modality is Enough: Multi-Modal Backdoor Attacks via Prototype-based Modality Rebalancing and Data Selection
Multi-modal deep learning has become the backbone of modern consumer electronics (CE), enabling seamless interaction in devices like smart security systems and health monitors. However, this also introduces new attack surface. While existing backdoor research focuses on uni-modal scenarios, successfully planting triggers on every modality is often blocked by the physical or protocol-driven constraint in practical CE environments. In response, we focus on a more challenging yet realistic threat model where the attacker has limited control and can only manipulate a single input modality to activate the backdoor. To address this, we introduce “prototype”, which represents each class in the embedding space, and propose a prototype-based modality rebalancing backdoor training method that stimulates specific modalities with poor backdoor learning performance. However, modality rebalancing inevitably introduces extra computation costs. Therefore, we reuse the computational elements of the prototype and propose a poisoning training data selection strategy based on the distance between the boundary samples group and poisoning samples, which can greatly save computational costs. Experimental results show that our attack success rate with a single weak modality’s trigger exceeds the latest multi-modal attack competitor by over 20% on average at the same poisoning rate.
- 19 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. PP (00983063) , 1
One Modality is Enough: Multi-Modal Backdoor Attacks via Prototype-based Modality Rebalancing and Data Selection
Multi-modal deep learning has become the backbone of modern consumer electronics (CE), enabling seamless interaction in devices like smart security systems and health monitors. However, this also introduces new attack surface. While existing backdoor research focuses on uni-modal scenarios, successfully planting triggers on every modality is often blocked by the physical or protocol-driven constraint in practical CE environments. In response, we focus on a more challenging yet realistic threat model where the attacker has limited control and can only manipulate a single input modality to activate the backdoor. To address this, we introduce “prototype”, which represents each class in the embedding space, and propose a prototype-based modality rebalancing backdoor training method that stimulates specific modalities with poor backdoor learning performance. However, modality rebalancing inevitably introduces extra computation costs. Therefore, we reuse the computational elements of the prototype and propose a poisoning training data selection strategy based on the distance between the boundary samples group and poisoning samples, which can greatly save computational costs. Experimental results show that our attack success rate with a single weak modality’s trigger exceeds the latest multi-modal attack competitor by over 20% on average at the same poisoning rate.
- 6 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. PP, 1-12
A Multiscale Collaborative Learning Method on Long-Term Time Series for Cyber-Physical-Social Intelligence
Cyber-physical-social system (CPSS) is an interactive system that combines digital technologies, physical infrastructures, and social dynamics. Within the CPSS framework, coordination of data flows, information flows, and knowledge flows facilitates seamless collaboration across human–machine boundaries. With the development and integration of artificial intelligence, cyber-physical-social intelligence (CPSI) has emerged, facilitating intelligent decision-making and promoting adaptive behavior in dynamic environments. CPSI aims to provide personal services through efficient CPSS data processing. Long-term time series (LTS), as one of the critical CPSS data, play a vital role in enabling CPSI to anticipate and address potential problems or changes in demand. However, real-world series commonly present irregular temporal patterns, which complicates the analysis of global information. Single-scale periodicity typically limits the exploration of temporal patterns, limiting the baseline’s ability to identify dynamic patterns. Moreover, LTS are frequently affected by noise, and excessively long sequences may hinder models from effectively capturing intersequence dependencies and the deep structural features embedded in the data. To address these challenges, this article proposes multiscale decomposition analysis baseline (MDAB), a highly effective long-term time series forecasting (LTSF) model. MDAB extracts multiscale components and performs collaborative learning of information across multiscales. For trend components, it integrates complex frequency-domain optimization to reduce noise, while for seasonal components, it employs efficient sparse attention to deeply analyze features and reduce memory usage. Comprehensive experiments carried out on several benchmark LTS datasets demonstrate that MDAB exceeds most existing LTSF models.
- 6 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. PP, 1-12
A Multiscale Collaborative Learning Method on Long-Term Time Series for Cyber-Physical-Social Intelligence
Cyber-physical-social system (CPSS) is an interactive system that combines digital technologies, physical infrastructures, and social dynamics. Within the CPSS framework, coordination of data flows, information flows, and knowledge flows facilitates seamless collaboration across human–machine boundaries. With the development and integration of artificial intelligence, cyber-physical-social intelligence (CPSI) has emerged, facilitating intelligent decision-making and promoting adaptive behavior in dynamic environments. CPSI aims to provide personal services through efficient CPSS data processing. Long-term time series (LTS), as one of the critical CPSS data, play a vital role in enabling CPSI to anticipate and address potential problems or changes in demand. However, real-world series commonly present irregular temporal patterns, which complicates the analysis of global information. Single-scale periodicity typically limits the exploration of temporal patterns, limiting the baseline’s ability to identify dynamic patterns. Moreover, LTS are frequently affected by noise, and excessively long sequences may hinder models from effectively capturing intersequence dependencies and the deep structural features embedded in the data. To address these challenges, this article proposes multiscale decomposition analysis baseline (MDAB), a highly effective long-term time series forecasting (LTSF) model. MDAB extracts multiscale components and performs collaborative learning of information across multiscales. For trend components, it integrates complex frequency-domain optimization to reduce noise, while for seasonal components, it employs efficient sparse attention to deeply analyze features and reduce memory usage. Comprehensive experiments carried out on several benchmark LTS datasets demonstrate that MDAB exceeds most existing LTSF models.