’Pre-training + fine-tuning’ has been widely used in various downstream tasks. Parameter-efficient fine-tuning (PEFT) has demonstrated higher efficiency and promising performance compared to traditional full-tuning. The widely used adapter-based and prompt-based methods in PEFT can be uniformly represented as adding an MLP structure to the pre-trained model. These methods are prone to over-fitting in downstream tasks, due to the difference in data scale and distribution. To address this issue, we propose a new adapter-based PEFT module, i.e., LoKi, which consists of an encoder, a learnable activation layer, and a decoder. To maintain the simplicity of LoKi, we use single-layer linear networks for the encoder and decoder, and for the learnable activation layer, we use a Kolmogorov-Arnold Network (KAN) with the minimal number of layers (only 2 KAN linear layers). With a bottleneck rate much lower than that of Adapter, LoKi is equipped with fewer parameters (only half of Adapter) and eliminates slow training speed and high memory usage of KAN. We conduct extensive experiments on LoKi under image classification and video action recognition across 9 datasets. LoKi demonstrates highly competitive generalization performance compared to other PEFT methods with fewer tunable parameters, ensuring both effectiveness and efficiency.

Autonomous vehicles (AVs) face significant threats to their safe operation in complex traffic environments. Adver-sarial policy for scenario generation has been established as a robust paradigm for enhancing AV resilience against adver-sarial perturbations through proactive exposure to synthetically engineered safety-critical scenarios. Training an attacker within an adversarial policy, allowing the target AV to expose vulnerabilities through interaction with this attacker. However, adversarial policies in existing methodologies often get stuck in a loop of over-exploiting established vulnerabilities, resulting in poor exploration for AVs. To overcome the limitations, we introduce a pioneering framework termed the vulnerability-aware and curiosity-driven adversarial reinforcement learning policy. Specifically, during the traffic vehicle attacker training phase, a surrogate network is employed to fit the value function of the AV victim, providing dense information about the victim's inherent vulnerabilities. Subsequently, random network distillation is used to characterize the novelty of the scenario, constructing an intrinsic reward to guide the attacker in exploring unexplored territories. Experimental results demonstrated that the adversarial policy embedded within the attacker exhibited robustness in convergence and significantly enhanced the activation of policy exposure in learning-based AVs, outperforming both other adversarial modalities and alternative reinforcement learning approaches, with a notable reduction in crash rates. The code is available at https://github.com/caixxuan/VCAT.

Existing autonomous driving systems are optimized for egocentric efficiency metrics, which are in fundamental conflict with the socialized expectations and habitual patterns of human drivers. This contradiction stems from the traditional approach's dual neglect of the trade-offs between self and other in driving decisions, and the culturally rooted social qualities of traffic interactions. To this end, this paper proposes a dual-adaptation framework that integrates social value orientation (SVO) and human-in-the-Ioop(HITL) guidance, modeling vehicular interactions as competitive-cooperative agents by quantifying the social utility function, and dynamically calibrating the SVO parameters with the help of real-time human feedback. The method innovatively transforms abstract social preferences into mathematically tractable decision boundaries, enabling the human-vehicle co-evolutionary mechanism to contextualize self-adaptation according to the regional driving etiquette, and thus cracking the inherent contradiction between individual trajectory optimization and group traffic harmony. Empirical studies based on Highway Env driving scenarios show that compared with pure reinforcement learning methods, this method reduces human-vehicle interaction conflicts while maintaining self-vehicle efficiency. The research results provide a quantifiable interaction paradigm and a verifiable training architecture for the construction of culturally-aware autonomous driving systems through the deep coupling of computational social value modeling and human social intelligence.

ABSTRACT Personalized federated learning represents a pivotal strategy for addressing the challenges posed by statistical heterogeneity in federated learning. Clients optimize their models by leveraging information from other clients through a global model. However, data heterogeneity constrains the generalization capacity of the global model, thereby degrading the feature representation capability of local client models, especially in clients with limited data. In response to this challenge, we propose the Federal Merging Subgroup Information (FedMSI) method to augment personalized information in personalized federated learning. On the server side, FedMSI employs model clustering to identify subgroups of clients with similar personalized data distributions. It then aggregates cluster center models within each subgroup and transmits them to clients for use in the subsequent round of assisted training. On the client side, FedMSI introduces a local optimization objective that incorporates the cluster center model, enabling the extraction of informative knowledge to enhance local training. Experiments demonstrate the effectiveness of FedMSI across different datasets, data heterogeneity levels, and data sizes. Ablation experiments further confirm the effectiveness of the design of the local optimization objective. Compared to state‐of‐the‐art methods, FedMSI achieves a 13.16% improvement in scalability performance accuracy.

Abstract With oceanic reverberation and a large amount of data being the main sources of interference for underwater acoustic target detection, it is difficult to obtain a more robust detection performance by relying on the traditional constant false alarm rate (CFAR) detection method. An adaptive sonar CFAR detection method based on a back propagation (BP) neural network is proposed. The method combines the artificial intelligence algorithm and the traditional detection algorithm, and uses the classification ability of the algorithm to select the detection algorithm, which can effectively improve the adaptation ability of the algorithm and the environment and the false alarm control ability. The method combines the artificial intelligence algorithm and the traditional detection algorithm, and uses the classification ability of the algorithm to select the detection algorithm, which can effectively improve the adaptation ability of the algorithm and the environment and the false alarm control ability. This method uses a BP neural network to train the target echo signal to complete the clutter background classification and establish the clutter background recognition classification set. According to the output result of each classification, the best CFAR detector is selected from four CA/SO/GO/OS‐CFAR detectors to detect the target. The simulation results show the detection performance of the proposed method in a uniform environment, a multi‐target environment, and a clutter edge environment. The results show that the environment adaptability is strong for different clutter backgrounds, which further improves the control ability of false alarms under a non‐uniform background.