The Internet of Vehicles (IoV) is rapidly evolving, but it also faces significant security threats, including Denial of Service (DoS) attacks and deceptive behaviors. The construction of an intrusion detection framework for the IoV encounters the following challenges: 1) the pervasive Non-IID (Non-Independent and Identically Distributed) issue within datasets and 2) in existing applications combining Federated Learning (FL) with blockchain, node load issues are often overlooked. To address these challenges, we propose a novel scheme that integrates Federated Distillation (FD) with blockchain technology to safeguard the IoV from network attacks. First, the scheme combines Convolutional Neural Networks (CNN) and FD for intrusion detection, which both preserves the privacy of vehicle users’ data and effectively shares knowledge. Next, for each FD client, we construct a selector based on density ratio estimation to filter out accurate and reliable knowledge from local predictions, thereby effectively mitigating the performance degradation caused by the Non-IID data problem. Then, the DQN (Deep Q-Network) integrated PBFT algorithm (DQN-PBFT) is proposed to reduce the number of consensus nodes using the DQN algorithm, enhancing load balancing and fairness during the FD and block generation processes. Finally, experimental results show that, in a Non-IID environment, the proposed scheme outperforms the baseline scheme by approximately 30% in accuracy. Compared to other consensus algorithms, DQN-PBFT demonstrates greater advantages in FD environments, exhibiting higher consensus efficiency.

In the Internet of Vehicles (IoV), the large number of computation tasks generated by intelligent vehicles can usually be decomposed into multiple subtasks with timing and data dependencies. Due to the limited computation resources of the vehicle itself, effectively handling these subtasks becomes a significant challenge. In this paper, task processing delay is modelled by using a directed acyclic graph (DAG) to represent the dependencies between subtasks and the longest path as the total delay of the task. Secondly, this paper introduces the concept of latest start time (LST) to measure the urgency of subtasks. Based on this, we design a LST-based computation resource allocation (LST-CRA) algorithm, which allows vehicular edge computing (VEC) server to allocate computation resources efficiently according to the two key characteristics of subtasks, i.e., time constraint and data volume. Furthermore, in order to optimize the offloading process of subtasks, we model the problem as a multi-agent collaborative optimization decision-making problem. By employing the multi-agent deep deterministic policy gradient (MADDPG) algorithm, we propose the LST-CRA driven MADDPG-based computation offloading (LDMCO) algorithm to meet the optimization requirements of subtask offloading. Experimental results show that the LDMCO algorithm and the LST-CRA algorithm demonstrate significant advantages in terms of task processing delay and task completion ratio. Even under high loads or strict time constraints, these algorithms demonstrate strong robustness and balanced system performance.

Novel view synthesis aims to generate new view images of a given view image collection. Recent attempts address this problem relying on 3D geometry priors (e.g., shapes, sizes, and positions) learned from multi-view images. However, such methods encounter the following limitations: 1) they require a set of multi-view images as training data for a specific scene (e.g., face, car or chair), which is often un-available in many real-world scenarios; 2) they fail to ex-tract the geometry priors from single-view images due to the lack of multi-view supervision. In this paper, we pro-pose a Geometry-enhanced NeRF (G-NeRF), which seeks to enhance the geometry priors by a geometry-guided multi-view synthesis approach, followed by a depth-aware training. In the synthesis process, inspired that existing 3D GAN models can unconditionally synthesize high-fidelity multi-view images, we seek to adopt off-the-shelf 3D GAN models, such as EG3D, as a free source to provide geometry priors through synthesizing multi-view data. Simultaneously, to further improve the geometry quality of the synthetic data, we introduce a truncation method to effectively sample la-tent codes within 3D GAN models. To tackle the absence of multi-view supervision for single-view images, we design the depth-aware training approach, incorporating a depth-aware discriminator to guide geometry priors through depth maps. Experiments demonstrate the effectiveness of our method in terms of both qualitative and quantitative results.