Selected work

Representative Papers

Secure Cooperative THz ISAC via Mamba Empowered Graph Neural Network Precoding

Aug 11, 2026

This work addresses the dual challenges of eavesdropping by malicious targets and high-precision localization in cooperative terahertz OFDM bistatic integrated sensing and communication systems. To this end, the authors propose a novel joint optimization framework that simultaneously designs analog beamforming, digital precoding, true time delays, and the covariance matrix of sensing signals. The approach innovatively integrates the Mamba state space model with graph neural networks, leveraging heterogeneous graph representations to capture interactions among users, targets, and base stations, while exploiting Mamba’s dynamic selection mechanism for context-aware optimization. By incorporating near-field channel modeling and true-time-delay beamforming, the proposed method significantly outperforms existing baselines in terms of secrecy rate, computational efficiency, and generalization capability, effectively achieving a balance between secure communication and high-accuracy sensing.

0 citationsRead paper

Intelligent Wiretap Code Design: Exploiting Wireless Endogenous Security via Information Theory and Deep Learning Integration

Aug 10, 2026

This work addresses the reliance on traditional cryptographic assumptions by proposing a wiretap coding scheme within a semantic communication framework that leverages the intrinsic randomness of wireless channels to jointly ensure security and reliable transmission. The design employs mutual information (MI) and generalized mutual information (GMI) as optimization criteria for two canonical eavesdropping scenarios, respectively, and integrates maximum a posteriori (MAP) decoding with deep learning–driven discrete semantic representations. By uniquely unifying semantic communication, information-theoretic metrics, and deep learning for wiretap code construction, this approach significantly reduces information leakage to eavesdroppers of varying capabilities while maintaining high reliability for the legitimate receiver.

0 citationsRead paper

Attention-Only White-Box Transformer via LeJEPA-Based Self-Supervised Pretraining

Aug 04, 2026

This work addresses a key limitation in existing self-supervised learning approaches for white-box networks, which often decouple architectural derivation from learning objectives. The paper proposes the first unified optimization framework that jointly integrates LeJEPA-based self-supervised learning with white-box Transformer design. By solving a sparse rate-distortion objective via the Alternating Direction Method of Multipliers (ADMM), the method yields a pure attention architecture composed solely of attention mechanisms, achieving the first MLP-free white-box Vision Transformer (ViT). This finding reveals redundancy in the standard ViT’s MLP blocks. The resulting model attains competitive performance—88.88% and 63.54% accuracy on CIFAR-10 and CIFAR-100, respectively—while reducing parameter count by 31% relative to the baseline. Notably, even after removing the MLP entirely, yielding a 66% model compression, the architecture maintains strong performance.

0 citationsRead paper

DM-KG: A Novel Method for Boosting Spatial Cognition of Vision-Language Models in Street View Imagery

Jul 13, 2026

This work addresses the issue of spatial-semantic hallucination in vision-language models when processing street-view images by introducing the DM-KG framework, which, for the first time, injects a directional-metric knowledge graph as geometric prior to enhance spatial reasoning. The method leverages panoptic segmentation and metric depth estimation to derive entity-level 3D coordinates, encoding pairwise clock-face azimuths and Euclidean distances between entities into a structured knowledge graph. Experimental results on public spatial question-answering benchmarks demonstrate that the proposed approach reduces the mean absolute error in distance estimation by 31.1% and decreases the mean angular error in direction judgment by 65.8%, while maintaining high question-answering success rates, thereby significantly improving both the accuracy and interpretability of spatial reasoning.

0 citationsRead paper

Physics-Informed Structure Anchoring With Capture-Aware Prototype Calibration for Cross-Environment RF Fingerprinting

Jul 06, 2026

This work addresses the significant performance degradation of multi-antenna RF fingerprinting when deployed across diverse environments, primarily caused by discrepancies in receiver array topology, dynamic carrier frequency offsets (CFO), and capture-dependent variations. To mitigate these challenges, the authors propose the PISA-CAPC framework, which uniquely integrates array topology graphs with CFO dynamics to construct physics-informed structural anchors. Furthermore, they introduce an unsupervised Capture-Aware Prototype Calibration (U-CAPC) mechanism that operates without target-domain labels or updates to the backbone network, effectively decoupling source-domain representation learning from target-domain decision calibration. Evaluated on a real-world multi-antenna Wi-Fi dataset comprising ten transmitters, the method achieves an average Macro-F1 score of 0.9257 under balanced conducted settings, demonstrating the complementary benefits and efficacy of its constituent components.

0 citationsRead paper

Recent publications

Latest Papers

Secure Cooperative THz ISAC via Mamba Empowered Graph Neural Network Precoding

Aug 11, 2026

This work addresses the dual challenges of eavesdropping by malicious targets and high-precision localization in cooperative terahertz OFDM bistatic integrated sensing and communication systems. To this end, the authors propose a novel joint optimization framework that simultaneously designs analog beamforming, digital precoding, true time delays, and the covariance matrix of sensing signals. The approach innovatively integrates the Mamba state space model with graph neural networks, leveraging heterogeneous graph representations to capture interactions among users, targets, and base stations, while exploiting Mamba’s dynamic selection mechanism for context-aware optimization. By incorporating near-field channel modeling and true-time-delay beamforming, the proposed method significantly outperforms existing baselines in terms of secrecy rate, computational efficiency, and generalization capability, effectively achieving a balance between secure communication and high-accuracy sensing.

0 citationsRead paper

Intelligent Wiretap Code Design: Exploiting Wireless Endogenous Security via Information Theory and Deep Learning Integration

Aug 10, 2026

This work addresses the reliance on traditional cryptographic assumptions by proposing a wiretap coding scheme within a semantic communication framework that leverages the intrinsic randomness of wireless channels to jointly ensure security and reliable transmission. The design employs mutual information (MI) and generalized mutual information (GMI) as optimization criteria for two canonical eavesdropping scenarios, respectively, and integrates maximum a posteriori (MAP) decoding with deep learning–driven discrete semantic representations. By uniquely unifying semantic communication, information-theoretic metrics, and deep learning for wiretap code construction, this approach significantly reduces information leakage to eavesdroppers of varying capabilities while maintaining high reliability for the legitimate receiver.

0 citationsRead paper

Attention-Only White-Box Transformer via LeJEPA-Based Self-Supervised Pretraining

Aug 04, 2026

This work addresses a key limitation in existing self-supervised learning approaches for white-box networks, which often decouple architectural derivation from learning objectives. The paper proposes the first unified optimization framework that jointly integrates LeJEPA-based self-supervised learning with white-box Transformer design. By solving a sparse rate-distortion objective via the Alternating Direction Method of Multipliers (ADMM), the method yields a pure attention architecture composed solely of attention mechanisms, achieving the first MLP-free white-box Vision Transformer (ViT). This finding reveals redundancy in the standard ViT’s MLP blocks. The resulting model attains competitive performance—88.88% and 63.54% accuracy on CIFAR-10 and CIFAR-100, respectively—while reducing parameter count by 31% relative to the baseline. Notably, even after removing the MLP entirely, yielding a 66% model compression, the architecture maintains strong performance.

0 citationsRead paper

DM-KG: A Novel Method for Boosting Spatial Cognition of Vision-Language Models in Street View Imagery

Jul 13, 2026

This work addresses the issue of spatial-semantic hallucination in vision-language models when processing street-view images by introducing the DM-KG framework, which, for the first time, injects a directional-metric knowledge graph as geometric prior to enhance spatial reasoning. The method leverages panoptic segmentation and metric depth estimation to derive entity-level 3D coordinates, encoding pairwise clock-face azimuths and Euclidean distances between entities into a structured knowledge graph. Experimental results on public spatial question-answering benchmarks demonstrate that the proposed approach reduces the mean absolute error in distance estimation by 31.1% and decreases the mean angular error in direction judgment by 65.8%, while maintaining high question-answering success rates, thereby significantly improving both the accuracy and interpretability of spatial reasoning.

0 citationsRead paper

Physics-Informed Structure Anchoring With Capture-Aware Prototype Calibration for Cross-Environment RF Fingerprinting

Jul 06, 2026

This work addresses the significant performance degradation of multi-antenna RF fingerprinting when deployed across diverse environments, primarily caused by discrepancies in receiver array topology, dynamic carrier frequency offsets (CFO), and capture-dependent variations. To mitigate these challenges, the authors propose the PISA-CAPC framework, which uniquely integrates array topology graphs with CFO dynamics to construct physics-informed structural anchors. Furthermore, they introduce an unsupervised Capture-Aware Prototype Calibration (U-CAPC) mechanism that operates without target-domain labels or updates to the backbone network, effectively decoupling source-domain representation learning from target-domain decision calibration. Evaluated on a real-world multi-antenna Wi-Fi dataset comprising ten transmitters, the method achieves an average Macro-F1 score of 0.9257 under balanced conducted settings, demonstrating the complementary benefits and efficacy of its constituent components.

0 citationsRead paper