As autonomous vehicles gain popularity, ensuring their safety and security becomes important. Multi-agent Collaborative Perception (MACP) enhances autonomous cars' perception capabilities. Despite its benefits, MACP is vulnera-ble to adversarial attacks, wherein malicious agents introduce adversarial examples to compromise the perception of recipient vehicles. This paper investigates the vulnerability of MACP to adversarial attacks, analyzing both white-box (where the attacker has access to the target model's gradients) and black-box scenarios (where the attacker lacks such access). We introduce an Agent-Aware Defense System designed to minimize these vulnerabilities in generic scenario. Through the application of the Fast Gradient Signed Method and Transferability techniques, we generate perturbations to assess the robustness of MACP models. Our results indicate that white-box attacks are more effective than black-box attacks and that early-stage collaborative methods are particularly vulnerable to adversarial manipulation. This study shows that integrating robust defense mechanisms in the development of collaborative perception frameworks for autonomous vehicles is important.
The general goal of traffic modeling is to predict the size of traffic flows in individual roads of the modeled road network. A commonly employed technique is the user-equilibrium (UE) traffic assignment (TA). There are multiple algorithms for solving TA with different convergence speeds. One of the fastest algorithms is the Improved Traffic Assignment by Paired Alternative Segments (iTAPAS) algorithm. In this paper, an optimized implementation of the iTAP AS algorithm is described and compared to its original implementation and to the sequential and parallel versions of the B algorithm, which we implemented during our previous research. An important part of the paper is the discussion of the parallelization of the iTAP AS algorithm, which is our ultimate goal.
This paper presents a simulation-based comparison of deep learning and reinforcement learning (RL) architectures for automatic modulation classification (AMC) under low-SNR conditions and real-time constraints. We benchmark a robust convolutional neural network (CNN), propose a lightweight dilated convolutional network (DCN-MODNET), and evaluate an exploratory RL agent using a custom Gym simulation. Our goal is to assess trade-offs between accuracy, inference latency, and computational cost for deployment in cognitive radio networks (CRNs). Results show that the DCN-MODNET offers strong performance with reduced complexity, while the RL agent shows niche utility in ultra-low SNRs but remains unsuitable for real-time use.
Artificial intelligence is transforming science into a highly automated process, accelerating the pace of scientific discovery. We argue that automatically adapting simulation models to new data, knowledge about mechanisms, or research questions plays a key role in this change of paradigm. To this end, we explore the use of the MAPE- K (Monitor, Analyze, Plan, and Execute, over a shared Knowledge base) framework from self-adaptive software systems to realize a self-adaptive simulation model. Such a model will preserve its currency, reveal new insights, and provide enhanced predictions as well as effective feedback to the observed system. As a case study, we use a cell-biological model of glucose-stimulated insulin secretion. We conclude with a discussion relating our approach to digital twins and current developments in inferring simulation models from data that opens up questions to be pursued in future research.
Predicting how multiple road users move is crucial for autonomous driving, but accuracy, reliability, and clarity are often hard to balance. This paper reviews hybrid methods that combine kinematic, map, and behavioral information with data-driven models. It organizes hybrid methods by where constraints are applied and compares them with classical and deep models using standard datasets and metrics in a compact table. It then discusses three key deployment challenges, which are handling domain shifts, calibrating uncertainty and interpretability, and mitigating closed-loop degradation-and outline practical checks for integrating hybrids into planning pipelines. Its goal is to provide a concise, practice-oriented perspective that shows when hybrids improve reliability and where they add complexity, to guide the design of accurate and trustworthy trajectory predictors.
