Building trustworthy intelligent systems

Multimodal 3D object detection is sensitive to feature misalignment and biased instance selection when point clouds are voxelized and projected onto the bird's-eye-view (BEV) plane. This problem mainly arises because voxel discretization maps continuous coordinate distributions to discrete indices, thereby reducing the sub-voxel geometric fidelity of multimodal features. To address this issue, we propose a lightweight extension of the IS-Fusion baseline by introducing continuous geometric priors. The proposed high-fidelity continuous coordinate encoding (HCE) preserves the continuous coordinate centroid within each voxel. Foreground-aware geometric enhancement (FGE) uses occupancy density, local geometric saliency, and normalized coordinate priors to produce residual gates that strengthen likely foreground regions. Geometry-guided instance fusion (GGI) further uses continuous geometric scores to conservatively calibrate semantic heatmap-based candidate ranking and instance query features. Experiments on the nuScenes validation set show that the proposed method improves overall detection performance, achieving 72.58 mAP and 73.83 NDS, while consistently reducing geometry-related errors, including mATE, mASE, and mAOE. These results suggest that continuous geometric priors are effective and can be integrated into multimodal 3D detection frameworks without disrupting the original fusion pipeline.

To address the security challenges such as data theft, information tampering and physical attacks faced by lightweight devices in the Internet of Vehicles (IoV), this paper proposes a lightweight two-way authentication scheme PUF-ILM that integrates Physical Unclonable Function (PUF) and Improved Logistic Map (ILM). The PUF generates unique challenge-response pairs (CRPs) for each device using its inherent physical randomness, thereby achieving reliable device identity authentication; the ILM encrypts and obfuscates the transmitted CRPs by enhancing the chaotic characteristics, expanding the parameter range and eliminating the periodic window, effectively resisting modeling analysis and eavesdropping attacks. Security analysis indicates that this scheme possesses 11 complete security attributes, including anti-cloning attack, anonymity, two-way authentication, forward/backward security and anti-machine learning modeling attack. Performance evaluation shows that PUF-ILM maintains a low communication overhead while having a simple system structure, does not rely on complex hardware or external trusted institutions, and combines high security and strong practicality, especially suitable for large-scale deployment in resource-constrained Internet of Vehicles environments.

Accurate and robust 3D object detection is essential for autonomous driving perception. Although LiDAR and cameras provide complementary geometric and semantic information, existing fusion methods still face two major challenges: insufficient representation of image features before fusion and limited modeling of global cross-modal relationships during fusion. To address these issues, this paper proposes a dual-module collaborative LiDAR-camera fusion framework for 3D object detection. The key novelty of the proposed method is not simply combining enhancement modules, but systematically treating image feature augmentation and global cross-modal fusion as two coupled bottlenecks. By improving visual representations before fusion and strengthening global associations between heterogeneous features during fusion, the proposed framework enables more effective multimodal interaction in complex driving scenarios. Experiments on the nuScenes dataset demonstrate that the proposed method achieves 69.3% mAP and 72.9% NDS, outperforming representative single modal and multimodal detection methods. In addition, the model maintains a compact scale with 36.8M parameters, 429.6G FLOPs, and 4.44 FPS. These results show that the proposed method provides an effective and computationally acceptable solution for robust LiDAR-camera 3D object detection.

HARD-V2X addresses hard-target evidence preservation and recovery in cooperative LiDAR-based V2X 3D detection. Distant objects, sparse-point objects, and easily suppressed traffic categories often suffer from weak local evidence, unstable responses, or unfavorable candidate competition. Although cooperative LiDAR perception can introduce complementary observations from multiple agents, useful hard-target evidence may still be insufficiently aligned, prematurely suppressed, or incompletely recovered during cooperative processing. We therefore propose HARD-V2X, a cooperative detection framework that improves this process through three coupled components: Neighborhood-Aware Heterogeneous Alignment preserves aligned cross-agent evidence under residual pose errors, Class-Aware Regional Proposal introduces class-aware candidate reservation under a fixed candidate-level interaction budget, and Candidate-Guided Cross-Scale Interaction recovers incomplete evidence through uncertainty-guided contextual compensation. On the real-world V2X-Real dataset, HARD-V2X achieves 0.617 mAP@0.3 and 0.524 mAP@0.5, outperforming representative V2X detection baselines. Additional distance-related, sparsity-related, and category-sensitive analyses further support the effectiveness of evidence preservation and recovery for robust V2X 3D detection.

The existing network authentication mechanisms in dynamic Internet of Vehicles (V2X) environments suffer from high communication overhead and inadequate security guarantees. To address these limitations, we propose PACLA, a lightweight mutual authentication protocol specifically engineered for resource-constrained V2X nodes. Our protocol integrates SoftPUF and the NIST-standardized lightweight cipher Ascon within a cross-layer collaboration architecture. SoftPUF is a software-based approach that emulates physical unclonable function (PUF) behavior. It overcomes the hardware dependency of traditional PUF and allows secure and efficient authentication for a wide range of devices in V2X networks. This synergy facilitates the creation of device-specific fingerprints for each session while ensuring efficient authenticated encryption with associated data (AEAD) and integrated key derivation via Ascon's XOF mode. PACLA uses a dynamic challenge-response mechanism and dual unencrypted binding to provide forward confidentiality and effectively resist replay attacks and machine learning modeling attacks. We evaluate the performance of the proposed scheme in terms of computational overhead and communication overhead. In addition, we conduct formal security analysis using the Real-Or-Random (ROR) model and the Scyther verification tool. The results demonstrate that our scheme achieves significant advantages in both security and performance.

With the rapid development of 6G networks, network security issues have become increasingly prominent, especially the security of identity authentication has become a key challenge. Traditional identity authentication methods face many limitations in the 6G environment, including vulnerability to cloning attacks and key leakage. This paper proposes a Mobile USIM identity authentication scheme based on a physical unclonable function (PUF), leveraging the hardware uniqueness and unclonability of PUFs to enhance the security and reliability of identity authentication. Through theoretical analysis and experimental verification, this paper proves the feasibility and superiority of the scheme in 6G networks. Experimental results show that the improved ring oscillator (RO) PUF in the scheme achieves an average inter-chip Hamming Distance (HD) of approximately 50% for 128-bit and 256-bit response lengths, demonstrating excellent uniqueness; under 2% bit-flipping noise (simulating environmental interference), the PUF maintains an average intra-chip reliability of 98% for 128-bit and 512-bit response lengths with a small standard deviation, ensuring stable output. The proposed PUF-based authentication scheme not only effectively resists cloning, replay, and physical attacks, but also features low resource consumption and high authentication efficiency, with lightweight hash and XOR operations as core algorithms. Formal security analysis via Python-based symbolic validation (max depth=6, 12s time limit) verifies no session key leakage in passive attack, replay and fake base station scenarios within the test bounds. This study provides a new solution for identity authentication in 6G networks, which has important theoretical significance and practical application value.