Sparse-view tomographic reconstruction aims to recover 3D volumes from limited projection views, but often suffers from incomplete structures and volumetric artifacts. Gaussian splatting has recently emerged as an efficient representation for continuous volumetric modeling, reducing memory cost compared to voxel grids and training time compared to implicit methods. However, existing Gaussian splatting methods for CT reconstruction struggle with needle-like artifacts in sparse-view settings. To address this, we introduce two key contributions. First, we propose a structure-aware initialization strategy that uses gradient and density magnitude from preliminary reconstructions to intelligently place Gaussian primitives in high-contrast regions. Second, we adapt the well-established Beer-Lambert law from CT physics to stabilize Gaussian splatting optimization, transforming the exponential attenuation relationship into a linear domain that mitigates vanishing gradients, and stabilizes optimization. Together, these innovations lead to sharper and more stable reconstructions, achieving average improvements of 2.32 % in PSNR and 2.41 % in SSIM while using 6.47 % fewer primitives across three standard CT datasets.

Large language models (LLMs) deployed on edge devices (smart-phones, IoT) require small parameters for efficiency, but state-of-the-art prompt injection defenses like SecAlign only work effectively at 8B+ parameters, leaving edge deployments vulnerable. We address this gap by transferring SecAlign's defense properties to 1B edge LLMs through on-policy distillation with reinforcement learning (RL). During on-policy distillation, the student models learn from their own failure modes and is inexpensive to train and evaluate rather than imitating teacher generated responses. The student generates responses to prompt-injected inputs and learns from teacher-guided preferences using importance sampling with negative reverse KL divergence as an advantage function. Through experiments on 1B Llama model, our on-policy RL distillation approach dramatically outperforms standard knowledge distillation with reduced attack success rate, and achieves comparable performance to 8B SecAlign teacher model with Llama backbone, enabling effective compression of state-of-the-art defenses to edge LLMs.

Graph Neural Networks (GNNs) power applications such as content recommendation, knowledge graph reasoning, and social network analysis, where modeling both structure and features is essential.Knowledge Distillation (KD) enables transferring knowledge from large GNNs to compact models for efficient deployment, yet most approaches rely on a pre-trained teacher.We propose a mutual learning framework in which shallow GNNs collaboratively distill knowledge by iteratively exchanging predictions during training.The framework integrates adaptive logit weighting to balance peer influence and entropy enhancement to promote exploration and prevent early convergence.Experiments on multiple benchmark datasets show that our approach improves GNN performance and that the learned knowledge can be effectively transferred to lightweight graph-less models, offering a scalable alternative for graph learning.

Graph Neural Networks (GNNs) are deep learning models designed to address the complexities of graph-structured, non-Euclidean data. Due to their complexity, knowledge distillation (KD) is often employed to transfer knowledge from a GNN to a simpler, more efficient student model, such as a Multi-Layer Perceptron (MLP), enabling deployment in large-scale industrial applications. However, KD can inadvertently leak sensitive information from the teacher to the student, posing significant privacy risks. We present the first membership inference attacks targeting GNNs in KD pipeline, showing that student MLPs can reveal whether a node appeared in the teacher’s training data. Our attacks operate in a black-box setting, requiring access only to the student outputs, and remain effective in cross-dataset scenarios. Experimental evaluations across four GNN models and eight datasets show the effectiveness of our approach, achieving up to 0.9014 precision under low FPR of 1% in cross-dataset settings. These results expose significant vulnerabilities in GNN-based KD frameworks, emphasizing the need for strong security measures during the KD process involving GNNs.

Vision-language models (VLMs) demonstrate impressive zero-shot and few-shot learning capabilities, making them essential for several downstream tasks. However, fine-tuning these models at scale remains challenging, particularly in federated environments where data is decentralized and non-iid across clients. Existing parameter-efficient tuning methods like LoRA (Low-Rank Adaptation) reduce computational overhead but struggle with heterogeneous client data, leading to suboptimal generalization. To address these challenges, we propose FedVLM, a federated LoRA fine-tuning framework that enables decentralized adaptation of VLMs while preserving model privacy and reducing reliance on centralized training. To further tackle data heterogeneity, we introduce personalized LoRA (pLoRA) which dynamically adapts LoRA parameters to each client’s unique data distribution, significantly improving local adaptation while maintaining global model aggregation. Experiments on the RLAIF-V dataset show that pLoRA improves client-specific performance by 24.5% over standard LoRA, demonstrating superior adaptation in non-iid settings. FedVLM provides a scalable and efficient solution for fine-tuning VLMs in federated settings, advancing personalized adaptation in distributed learning scenarios.

The Internet of Vehicles (IoV) is envisioned to improve road safety, reduce traffic congestion, and minimize pollution. However, the connectedness of IoV entities increases the risk of cyber attacks, which can have serious consequences. Traditional intrusion detection systems (IDS) transfer large amounts of raw data to central servers, leading to potential privacy concerns. Also, training IDS on resource-constrained IoV devices generally can result in slower training times and poor service quality. To address these issues, we propose a split learning-based privacy-preserving IDS that deploys IDS on edge devices without sharing sensitive raw data. In addition, we propose a regret minimization-based adaptive offloading technique that reduces the training time on resource-constrained devices. Our approach effectively detects anomalous behavior while preserving data privacy and reducing training time, making it a practical solution for IoV. Experimental results show the effectiveness of our approach and its potential to enhance the security of the IoV network.

The Internet of Vehicles (IoV) is an active area for innovation and an essential tool in achieving smart cities through the integration of vehicles with the Internet of Things (IoT). IoV is a distributed network that aids in handling the data generated by vehicular sensors and vehicle-to-everything communication (V2X), thus enabling novel applications such as autonomous driving and platooning while increasing safety and energy efficiency. In IoV, the sensors and the interdependent devices relay critical information for the efficient implementation of real-time applications in the ecosystem. Despite all these advancements, a vital challenge is establishing smooth communication among interconnected devices, concretely, interoperability in the IoV—a deceptively simple notion that is not yet fully addressed to achieve a fully integrated ecosystem. This is mainly because the networked domains, such as home, grid, and health care, are developed in silos, operating independently with diverse processes and protocols. Hence, seamless exchange of information is yet to be achieved across the ecosystem, hindering the maximization of the full promise of IoV. In this paper, we provide an in-depth analysis of the present state of interoperability and comprehensively survey the challenges in IoV. We present a taxonomy of interoperability approaches, review solutions that prior work have proposed, and provide insights on how to address the current challenges. Finally, we identify open problems that persist and future directions for research.