Nowadays, Unmanned Aerial Vehicle (UAV)-assisted Mobile Edge Computing (MEC) has been recognized as a promising technique for flexibly handling computation tasks in 5G advanced and 6G networks. This paper investigates the joint optimization of service caching and computation offloading within a dual time-scale framework. We maximize the caching utility and minimize the task processing delay by jointly optimizing service caching policies, UAV flight trajectory, and computation offloading decisions. Specifically, for the long-term problem, we use the Latent Dirichlet Allocation (LDA) model to predict user preferences, and propose a Lagrangian dual decomposition based algorithm. For the short-term problem, a self-attention based Multi-Agent Proximal Policy Optimization (MAPPO) algorithm is designed. Under the Centralized Training with Decentralized Execution (CTDE) framework, this algorithm integrates a multi-head self-attention mechanism with curriculum learning. Each UAV is regarded as an agent, and a self-attention encoder is integrated at the front-end of each Actor network. This enables the agent to dynamically capture the relative importance between itself and all users, and context-aware features are extracted to make more intelligent and trajectory designs. Through extensive simulation experiments, the long-term algorithm yields the performance improvements of 76.5% in cache hit rate and 66% in caching utility, as compared to the second best baseline. In dynamic scenarios, the short-term algorithm achieves a 16% reduction in total processing latency with respect to the proximal policy optimization policy.

The hypothesis of \textit{Universality} in interpretability suggests that different neural networks may converge to implement similar algorithms on similar tasks. In this work, we investigate two mainstream architectures for language modeling, namely Transformers and Mambas, to explore the extent of their mechanistic similarity. We propose to use Sparse Autoencoders (SAEs) to isolate interpretable features from these models and show that most features are similar in these two models. We also validate the correlation between feature similarity and~\univ. We then delve into the circuit-level analysis of Mamba models and find that the induction circuits in Mamba are structurally analogous to those in Transformers. We also identify a nuanced difference we call \emph{Off-by-One motif}: The information of one token is written into the SSM state in its next position. Whilst interaction between tokens in Transformers does not exhibit such trend.

In this paper, we propose a secure content caching and delivery strategy to support multi-level viewing services based on scalable video coding (SVC). With the existence of multiple eavesdroppers (Eves), the base layer (BL) and enhancement layers (ELs), that are encoded by SVC are cached in the macro-cell base station (MBS) and small-cell base stations (SBSs), respectively, using a probabilistic caching strategy. Power domain non-orthogonal transmission is utilized to deliver ELs for the serving SBS. We analyze the transmission secrecy performance for the end user. Specifically, employing the tool of stochastic geometry, the successful transmission probabilities (STPs) from the serving MBS and SBS are derived, and the worst-case secrecy outage probability (SOP) is also analyzed. After obtaining the expressions for STPs and SOP, a more comprehensive performance metric to evaluate the system security and caching efficiency, called the secrecy cache-aided data rate (SCADR), is analyzed and derived. Simulation results validate the accuracy of our derived STPs and SOP, and show how the SCADR varies with important system parameters under different caching schemes.

Reconfigurable Intelligent Surface (RIS) has emerged as a pivotal technology in the development of sixth-generation (6G) mobile communication systems. In this paper, a wireless communication system assisted by RIS in the air is studied, in which RIS is installed on the UAV as a mobile relay (RIS-UAV) to re-establish the line-of-sight (LoS) link between the base station and users. Considering the active beamforming vector of base station, UAV trajectory and phase shift of RIS, a joint optimization scheme is proposed to maximize the energy efficiency of the system. Aiming at this optimization problem, a dual DDQN structure algorithm (PER-TDDQN) based on priority experience replay mechanism is proposed to solve this multi-objective joint optimization problem. The simulation results show that the proposed algorithm can effectively improve the energy efficiency of the system.

This paper investigates a covert communication system enhanced by a reconfigurable intelligent surface (RIS), where an unmanned aerial vehicle (UAV) serves as the transmitter while a Warden moves freely within a certain range. To enhance both communication covertness and energy efficiency, we jointly optimize the UAV trajectory and RIS phase shifts to maximize system energy efficiency. The optimization problem is formulated as a Markov decision process (MDP), and a reinforcement learning-based approach, LSTM-DDQN, is proposed to exploit deep reinforcement learning (DRL) in dynamic environments. By integrating long short-term memory (LSTM) to capture temporal dependencies and double deep Q-network (DDQN) to mitigate Q-value overestimation, the proposed algorithm enhances stability and decision-making effectiveness. Simulation results demonstrate that LSTM-DDQN significantly improves energy efficiency while effectively reducing the risk of detection compared to conventional DDQN, validating its superiority in RIS-assisted UAV covert communication.

This paper proposes a novel framework for downlink multi-user cluster communications in non-orthogonal multiple access (NOMA) systems assisted by a simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS). We focus on the energy splitting (ES) protocol and account for practical issues such as hardware impairments (HIs) and error propagation from imperfect successive interference cancellation (I-SIC). The objective is to maximize the total user throughput by jointly optimizing user clustering, dynamic decoding order, base station active beamforming, and the STAR-RIS’s transmission and reflection beamforming. Furthermore, to address user mobility„ we design an efficient two-step algorithm: first, K-means is used for periodic user clustering, followed by proximal policy optimization (PPO), a deep reinforcement learning(DRL) algorithm, is applied to dynamically solve the aforementioned joint optimization problem. Simulation results show that the proposed algorithm significantly outperforms baseline DRL approaches in total throughput and that the STAR-RIS-NOMA system significantly exceeds conventional RIS-NOMA and OMA systems in performance.

To reduce the heavy traffic burden over backhaul links, this paper studies a new reconfigurable intelligent surface (RIS) aided wireless caching system. The RIS is equipped with a smart controller to control reflective elements. The functions of content caching and energy harvesting (EH) are endowed to the controller. A two-phase transmission policy is designed. In the first phase, the requested contents that are not locally cached are delivered by the base station (BS) to the RIS, and the smart controller harvests the energy from ambient radio frequency (RF) signals. In the second phase, the smart controller distributes cached contents utilizing non-orthogonal multiple access (NOMA) transmission with the aid of the harvested energy, while the RIS forwards the fetched contents. The beamforming vectors of the BS and the power allocation factors for NOMA transmission are jointly optimized to maximize the achievable sum rate of the system. The original problem is difficult to solve owing to coupled variables. After being processed by fractional programming (FP) and quadratic transformation (QT), it is divided into two sub-problems. The first sub-problem is solved by introducing auxiliary variables, and the second one is converted to be convex by rational transformations to the quality of service constraint. The optimal solution is achieved by alternatively solving the two sub-problems. Simulation results reveal that the original problem can converge quickly, and the proposed algorithm outperforms benchmarks in terms of the achievable sum rate.

This paper proposes a new joint random caching and hierarchical transmission scheme for delivering multimedia content in cache-assisted heterogeneous networks. We use scalable video coding (SVC) and wireless edge caching to offer personalized video-watching services to end users. Resorting to stochastic geometry, we derive new expressions for successful transmission probabilities (STPs). Using the derived STPs, the achievable transmission rates (ATRs) and transmission delay experienced by multimedia users are obtained. The Gaussian-Chebyshev quadrature is employed to approximate the STPs and ATRs in closed form to improve mathematical tractability. We also attain the user satisfaction index (USI) concerning user’s delay experience. A transmission delay minimization problem is formulated and solved efficiently using the standard gradient projection approach to determine the optimal layer placement design under a random caching criterion. Simulation results confirm the correctness of the derived STPs, show the negligible approximation loss of Gaussian-Chebyshev quadratures, and manifest the performance superiority of the devised SVC-based caching design to the energy-optimal caching and the non-caching approach.

This paper investigates the joint computation offloading and content caching scheme in fog radio access networks (F-RANs), where users and fog nodes (FNs) are all endowed with caching and computing capabilities. The Kuhn-Munkres (KM) algorithm is first used to make the user-and-FN matching decisions. We then comprehensively reveal the service modes when a user receives the requested data to perform a computation task. The optimization problem of minimizing the task response latency is formulated under the constraints of limited cache sizes for users and FNs, aiming to yield the optimal decisions for computation offloading and content caching. The formulated problem is NPhard, and transformed into the convex form by introducing the methods of difference of convex (DC) function, successive convex approximation (SCA), and arithmetic geometric mean (AGM) inequality. Simulation results demonstrate the fast convergence speed of the proposed algorithm, and show superior latency performance of our designed scheme to the benchmarks.

Wireless edge caching can alleviate the transmission pressure and improve the quality of experience (QoE) provided to users, which has been proposed as a promising technique in the sixth generation (6G) communication systems. This paper considers a clustered device-to-device (D2D) caching network, where users having similar request preference are classified into the same cluster. Due to the fact that recommendation policy will have a significant effect on user requests, the joint design of content recommendation and caching is investigated, aiming to enhancing the system caching efficiency (CE). Under the constraints of the recommendation quality, recommendation quantity and cache size, we formulate the CE maximization problem. The original problem is an NP-hard integer programming problem, making it difficult to solve. Therefore, it is partitioned into two sub-problems. The binary variables in the first sub-problem, namely, the recommendation design sub-problem, is relaxed to be continuous, and then the relaxed problem is solved by applying the methods of successive convex approximation and arithmetic geometric mean inequality. The recommendation decisions are recovered following a greedy strategy based principle. The second sub-problem, i.e., the caching design sub-problem, is addressed by using the smooth function to replace the $l_{0}$ norm. Simulation results demonstrate the fast convergence speed of the proposed algorithm, and show the obtained CE performance is superior to existing homogeneous recommendation method.

Confronted with the soaring growth of wireless traffic, edge caching has been deemed as a promising technique to address the challenges caused by redundant file deliveries and finite backhaul/fronthaul capacities. When making caching decisions, popularity prediction is preliminary but crucial. Con-sidering diversified viewing preferences of users, scalable video coding (SV C) is used to divide the file into multiple layers. The layers are used as caching units, and then delivered to users. Owing to this fact, we study the layer-based popularity prediction and content placement schemes. To protect users' privacy, a distributed framework based on federated learning is proposed to acquire the layer popularity. During the local training process, users are divided into privacy-sensitive and privacy-insensitive users. Taking users' privacy sensitivity into account, different types of nodes are nominated as edge nodes (EN s) for training, further reducing the risk of privacy leakage. The stacked auto encoder is used to extract features from both users and videos. By resorting to collaborative filtering, the features are used as inputs for predicting the layer popularity. Extensive experiments are conducted using a real dataset. Results reveal that, in contrast with the scheme that only provides basic viewing quality, the cache hit rate (CHR) is doubled by applying the SVC-based scheme when the cache size of each EN is set to 160 Mbits. We also find that the proposed algorithm improves the CHR by 55 % and decreases the transmission delay by 49 %, as compared to the Least Frequently Used (LFU) scheme.

Internet of Medical Things (IoMT) is regarded as an imperative technology for intelligent healthcare in the foreseeable 6G era. Due to the limited computing power of edge devices and task-related coupling, IoMT faces significant challenges. Considering the associative relationship among tasks, this paper proposes a computing offloading policy for multiple-user devices (UDs) under a multi-access edge computing (MEC) system. Specifically, we formulate the offloading scheme as a mixed-integer nonconvex optimization problem to minimize the total delay and energy consumption. To obtain the optimal solution, a double deep Q-network (Double DQN) based associative tasks computing offloading (DDATO) algorithm is then proposed, which can make the best offloading decision under the condition that tasks of UDs are associative. In addition, we use a dynamic ε−greedy strategy in the action selection section of the algorithm, thus preventing the algorithm from falling into a locally optimal solution. Simulation results demonstrate that compared with other existing methods, the proposed algorithm can lower the total cost more efficiently within the maximum delay and energy consumption tolerance.

Immersive virtual reality (VR) video is becoming increasingly popular owing to its enhanced immersive experience. To enjoy ultrahigh resolution immersive VR video with wireless user equipments, such as head-mounted displays (HMDs), ultralow-latency viewport rendering, and data transmission are the core prerequisites, which could not be achieved without a huge bandwidth and superior processing capabilities. Besides, potentially very high energy consumption at the HMD may impede the rapid development of wireless panoramic VR video. Multiaccess edge computing (MEC) has emerged as a promising technology to reduce both the task processing latency and the energy consumption for HMD, while bandwidth-rich terahertz (THz) communication is expected to enable ultrahigh-speed wireless data transmission. In this article, we propose to minimize the long-term energy consumption of a THz wireless access-based MEC system for high quality immersive VR video services support by jointly optimizing the viewport rendering offloading and downlink transmit power control. Considering the time-varying nature of wireless channel conditions, we propose a deep reinforcement learning-based approach to learn the optimal viewport rendering offloading and transmit power control policies and an asynchronous advantage actor-critic (A3C)-based joint optimization algorithm is proposed. The simulation results demonstrate that the proposed algorithm converges fast under different learning rates, and outperforms existing algorithms in terms of minimized energy consumption and maximized reward.