In recent years, tremendous success has been witnessed in Retrieval-Augmented Generation (RAG), widely used to enhance Large Language Models (LLMs) in domain-specific, knowledge-intensive, and privacy-sensitive tasks. However, attackers may steal those valuable RAGs and deploy or commercialize them, making it essential to detect Intellectual Property (IP) infringement. Most existing ownership protection solutions, such as watermarks, are designed for relational databases and texts. They cannot be directly applied to RAGs because relational database watermarks require white-box access to detect IP infringement, which is unrealistic for the knowledge base in RAGs. Meanwhile, post-processing by the adversary's deployed LLMs typically destructs text watermark information. To address those problems, we propose a novel black-box ''knowledge watermark'' approach, named RAG-WM, to detect IP infringement of RAGs. RAG-WM uses a multi-LLM interaction framework, comprising a Watermark Generator, Shadow LLM & RAG, and Watermark Discriminator, to create watermark texts based on watermark entity-relationship tuples and inject them into the target RAG. We evaluate RAG-WM across three domain-specific and two privacy-sensitive tasks on four benchmark LLMs. Experimental results show that RAG-WM effectively detects the stolen RAGs in various deployed LLMs. Furthermore, RAG-WM is robust against paraphrasing, unrelated content removal, knowledge insertion, and knowledge expansion attacks. Lastly, RAG-WM can also evade watermark detection approaches, highlighting its promising application in detecting IP infringement of RAG systems.

A typical supply chain consists of suppliers, manufacturers, distributors, and customers, with supply, production, and distribution being the key links. Coordinating and optimizing these stages to reduce waste and shorten delivery times poses a significant challenge. However, most existing research relies on heuristic algorithms and focuses primarily on the production and distribution stages. In complex large-scale scenarios, the computational efficiency and adaptability of heuristic algorithms often fall short. This paper investigates a combined scheduling problem and a heterogeneous vehicle routing problem. Unlike traditional heuristic approaches, we propose a novel deep-reinforcement learning model. Finally, a Multi-Rollout algorithm is employed for collaborative training. Experimental results demonstrate that the proposed algorithm delivers competitive performance compared to heuristic methods while significantly outperforming them in computation time for individual instances.

Deep reinforcement learning (DRL) has been widely applied to address inventory management problems. To tackle the challenges posed by factors such as backlog, multisource replenishment, and demand priority in multi-echelon inventory systems, this paper proposes a value decomposition-based multi-agent reinforcement learning (MARL) framework. The framework utilizes independent DQN networks, embedded with self-attention and GRU modules, to facilitate distributed learning of local action value functions, thus simplifying the complexity of the action space. Additionally, a hybrid network based on the multi-head attention mechanism is constructed to approximate the joint action value function, aiming to optimize the overall system cost. Experiments have been conducted on various types of supply chain networks, and the results demonstrate the effectiveness and scalability of the proposed framework.

Reliable and robust airport flight ground service resources collaborative scheduling has emerged as an issue of major concern at congested airports, which has great significance on airport service quality and operation efficiency. Due to the uncertainty of the service process of multiple vehicles in airport ground operations, the combination of simulation and optimization (sim-opt) techniques has been proven to be an effective means to solve the collaborative scheduling problem. However, existing models still have limitations on the reliability of feasible solutions and the robustness of evaluations. In this work, we developed a conflict resolution based sim-opt framework that combines the search capability of mathematical optimization models with the ability of simulation models to describe uncertainty. The dynamic window is introduced to divide the flight zone into multiple sub-intervals, which allows for iterative search and evaluation of the optimal schedule solution while reducing model calculation time. Additionally, a matching algorithm based on slack and tight constraints is proposed to optimize the possible resource requirements of each vehicle, thereby improving the initial schedule solution’s search depth and reliability. Furthermore, a conflict resolution feedback mechanism between adjacent windows is constructed to optimize the scheduling plan, reducing misjudgment and omission of the optimal solution and enhancing the sim-opt framework’s robustness. Finally, experiments on real airport datasets of three different scales demonstrate that our proposed method efficiently allocates resources with a smaller flight delay time as well as reducing total vehicle cost time during the flight ground service process.

Cross-domain wireless sensing has always been challenging due to the sensitivity of wireless signals to various environmental factors, which we refer to as subdomains. However, current efforts are limited to cross-one-subdomain tasks requiring target domain data for model training or multiple receivers for data collection. Taking common gesture recognition as an application example, we attempt to demonstrate the feasibility of cross-multiple-subdomain wireless sensing in the more challenging domain generalization (DG) setting. It is possible to extract domain-invariant features from one or several source domain(s), thereby avoiding the need for multiple receivers or target domain data. We also propose an intelligent wireless data augmentation technique based on subdomain-guided perturbations, named WiSGP. Specifically, the independent domain model generates perturbations in the direction of the largest subdomain variations. Then, these subdomain-guided perturbations augment the gesture model's input to enable better domain-invariant feature extraction, even when various subdomains interact. Similarly, gesture-guided perturbations augment the domain model's input, resulting in more accurate subdomain-guided perturbations and minimal gesture label changes. Extensive experiments have been conducted on three datasets collected from various NICs. In terms of room, location, orientation, and user subdomains, WiSGP exhibits excellent accuracy, generalizability, and portability for both cross-one-subdomain and cross-multiple-subdomain tasks.

Currently, available contactless continuous authentication (CA) techniques depend on physiological biometrics to identify individuals at long intervals through complex feature extraction, resulting in poor accuracy, high computation costs, and security vacuums during lengthy intervals. To address these issues, we propose WiPT, a WiFi-based contactless CA system that utilizes contextual features and behavioral biometrics to optimize contactless CA technology. Specifically, we designed a low-computation two-step user state detection (TUSD) mechanism that continuously monitors user states in real-time. It locks the system when the registered user leaves and allows user authentication only when the departing user returns. Therefore, it eliminates pointless periodic re-authentication and results in considerably shorter monitoring intervals while significantly reducing computation. Subsequently, benefiting from contextual features, WiPT can identify individuals using more detectable behavioral biometrics. We built a one-class classification model based on the Convolutional Autoencoder to automatically extract rich representations of WiFi signals associated with postural transition movements, resulting in lower authentication delay, higher accuracy, and anti-interference. WiPT was implemented by the widely available 802.11n devices and has been extensively evaluated with typical sit-to-stand postural transitions. WiPT achieves an average accuracy of 96.63% in authentication and 99.78% in defense across 30 subjects with an authentication delay of 5.59 milliseconds and a monitoring interval of 2 seconds. They are 4.87% and 5.03% more accurate and dozens of times less time-consuming than existing WiFi-based CA solutions.

The Internet of Things (IoT) has become more popular over the past decade. For the IoT to be successful, it is vital to track the location of these things (sensors or actuators). In this article, based on a new IoT underlying architecture, a narrowband-IoT (NB-IoT)-aided, bounding-based iterative and range-free method, named BBIL, is proposed to localize things. We make use of the location information of all anchor regular nodes that can help improve the localization accuracy as much as possible. Specifically, not only single-hop and multihop anchor things but also single-hop and multihop regular things are used for localization. In addition, the communication and computational loads of the local network are greatly decreased because the anchor things can directly access the Internet using the NB-IoT modules; hence, data can be sent to the Internet through a small number of hops in the local network and processed in the cloud/edge computing facilities in a centralized way. To balance the location accuracy and energy consumption of BBIL, we propose a theoretical model to obtain the optimal number of the anchor things. BBIL is evaluated and compared with the existing methods. The simulation results indicate that the average localization error of BBIL is less than 11.6%. Also, it performs well in anisotropic networks. In addition, we verified the validity of our method in real-world scenario.