Specific Emitter Identification (SEI) is crucial for ensuring the security of physical layer communication. However, signal characteristics can be affected by various factors such as environmental and equipment variations. An effective SEI system must continuously learn and adapt to these changes to maintain accurate signal recognition. This study proposes an advanced domain incremental learning (DIL) framework for SEI, named Adaptive Domain-Incremental Learning with Knowledge Replay and Domain Alignment (ADIRA). ADIRA employs knowledge replay and distillation strategies, along with adaptive coefficients, to balance the model’s performance in recognizing signals across both new and old domains. To address the variations in signal data feature distributions across different domains, we introduce a domain alignment strategy based on adversarial training. This approach integrates embedding distillation loss with supervised contrastive loss, significantly enhancing the model’s adaptability to domain changes. Experimental results on two benchmark datasets demonstrate that ADIRA achieves performance only 0.42% and 1.71% lower than joint training, with replay samples constituting just 1.1% and 1.5% of the training set, effectively mitigating catastrophic forgetting.

Deep reinforcement learning (DRL) has been successfully used for the joint routing and resource management in large-scale cognitive radio networks. However, it needs lots of interactions with the environment through trial and error, which results in large energy consumption and transmission delay. In this paper, an apprenticeship learning scheme is proposed for the energy-efficient cross-layer routing design. Firstly, to guarantee energy efficiency and compress huge action space, a novel concept called dynamic adjustment rating is introduced, which regulates transmit power efficiently with multi-level transition mechanism. On top of this, the Prioritized Memories Deep Q-learning from Demonstrations (PM-DQfD) is presented to speed up the convergence and reduce the memory occupation. Then the PM-DQfD is applied to the cross-layer routing design for power efficiency improvement and routing latency reduction. Simulation results confirm that the proposed method achieves higher energy efficiency, shorter routing latency and larger packet delivery ratio compared to traditional algorithms such as Cognitive Radio Q-routing (CRQ-routing), Prioritized Memories Deep Q-Network (PM-DQN), and Conjecture Based Multi-agent Q-learning Scheme (CBMQ).

Radio Frequency Fingerprint Identification (RFFI) has inspired numerous device authentication methods. However, the impact of receiver impairments on transmitter fingerprinting has not been adequately addressed in the existing literature. This letter proposes a receiver-agnostic RFFI model aimed at effectively extracting transmitter-related features from raw signals in the presence of receiver-related interference. Specifically, we employ distinct classifiers to initially extract transmitter-related and receiver-related features. The disentangled features from different samples are then cross-combined and used to obtain a transmitter classifier that is unaffected by receiver-related information. Additionally, a binary mask disentangler and a reconstruction module are incorporated to further extract transmitter-related information from the original samples. Experiments conducted on a real-world dataset demonstrate that our method outperforms four other comparison methods when processing data from unknown receivers.

Natural disasters often destroy critical infrastructure, such as terrestrial communication networks and transportation routes, thereby severely disrupting post-disaster rescue operations. To rapidly re-establish communication links and provide flexible computational support in disaster rescue scenarios, the integration of autonomous aerial vehicles (AAVs) and mobile edge computing (MEC) has emerged as a promising solution. Nevertheless, the highly complex and resource-constrained characteristics of disaster environments pose significant challenges for AAV-assisted computation task offloading. In this article, we investigate the joint task offloading and resource allocation (JTORA) problem to minimize the energy consumption associated with communication and computation during task offloading. Specifically, we develop a twin-delayed deep deterministic policy gradient (TD3)-based JTORA (JTORA-TD3) algorithm, which enables the AAV to optimize decisions of task offloading and resource allocation intelligently. To further enhance the training efficiency of the JTORA-TD3 algorithm in a complex disaster rescue environment, we integrate a large AI model (LAM) into the TD3 framework. Based on the textual interaction, we propose an LAM-enabled TD3-based JTORA (JTORA-LAM4TD3) algorithm. Simulation results demonstrate that the proposed JTORA-LAM4TD3 algorithm significantly outperforms baselines. These findings confirm the effectiveness of integrating LAMs with deep reinforcement learning (DRL) for solving the decision optimization problem.

Unmanned aerial vehicles (UAVs) have emerged as a promising solution for collaborative search missions in complex environments. However, in the presence of interference, communication disruptions between UAVs and ground control stations can severely degrade coordination efficiency, leading to prolonged search times and reduced mission success rates. To address these challenges, this paper proposes a novel multi-agent deep reinforcement learning (MADRL) framework for joint spectrum and search collaboration in multi-UAV systems. The core problem is formulated as a combinatorial optimization task that simultaneously optimizes channel selection and heading angles to minimize the total search time under dynamic interference conditions. Due to the NP-hard nature of this problem, we decompose it into two interconnected Markov decision processes (MDPs): a spectrum collaboration subproblem solved using a received signal strength indicator (RSSI)-aware multi-agent proximal policy optimization (MAPPO) algorithm and a search collaboration subproblem addressed through a target probability map (TPM)-guided MAPPO approach with an innovative action-masking mechanism. Extensive simulations demonstrate superior performance compared to baseline methods (IPPO, QMIX, and IQL). Extensive experimental results demonstrate significant performance advantages, including 68.7% and 146.2% higher throughput compared to QMIX and IQL, respectively, along with 16.7–48.3% reduction in search completion steps versus baseline methods, while maintaining robust operations under dynamic interference conditions. The framework exhibits strong resilience to communication disruptions while maintaining stable search performance, validating its practical applicability in real-world interference scenarios.

Radio frequency fingerprint identification (RFFI) is a widely used technique for authenticating equipment. It identifies transmitters by extracting hardware defects found in the RF front end. Recent research has focused on the impact of transmitters and wireless channels on radio frequency fingerprint (RFF). Most work is based on the same receiver assumption, while the influence of the receiver on RFF remains unresolved. This paper focuses on the impact of receiver hardware characteristics on RFF and proposes a few‐shot cross‐receiver RFFI method based on feature separation. Data augmentation with noise addition and simulated channels addresses sparse sample issues and enhances the model's robustness to channel variations. Simultaneously, feature separation is realized by reducing the correlation between transmitter and receiver features through classification loss and similarity loss. We evaluate the proposed approaches using a large‐scale WiFi dataset. It is shown that when a trained transmitter classifier is deployed on new receivers with only 30 samples per trained transmitter, the average identification accuracy of the proposed method is 83.6%. This accuracy is 9.45% higher than the baseline method without considering transmitter hardware influence. After fine‐tuning, the average identification accuracy can reach 98.25%.