With the rapid proliferation of wireless devices, ensuring secure communication over open wireless channels has become increasingly critical. Physical layer key generation has emerged as a lightweight and information-theoretically secure cryptographic mechanism, offering a promising complement to traditional cryptographic approaches. However, in Internet of Things (IoT) scenarios characterized by heterogeneous devices and highly dynamic environments, conventional physical-layer key generation methods suffer from low key generation rates, poor stability, and limited adaptability. To address these challenges, this paper proposes a novel AI-based physical layer key generation framework that integrates multi-source signal fusion and intelligent feature selection. Specifically, deep learning models are employed to fuse features from multiple heterogeneous wireless signal sources, enabling the extraction of high-quality randomness and enhancing key entropy and security. In parallel, an attention mechanism is employed to dynamically select the most suitable physical layer features based on real-time environmental conditions, thereby enhancing the system’s adaptability and robustness in complex IoT settings. Finally, we outline potential future research directions and discuss the feasibility of implementing the proposed framework in real-world deployments. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.

Federated Learning (FL) enables collaborative model training without sharing raw data, but remains vulnerable to poisoning attacks from malicious clients. Existing defenses are often reactive and require costly model retraining, making them inefficient and impractical for real-time protection. We propose FedCleaner, a server-side dual-mechanism framework that combines: Proactive Layer-Wise Anomaly Detection to identify poisoned updates in real time; Retroactive Contribution Erasure to efficiently unlearn malicious client influences without retraining. Experiments on datasets show that FedCleaner provides a scalable, privacy-preserving, and regulation-compliant solution to defend FL systems against persistent poisoning threats. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.

Tool-using LLM agents for health monitoring raise critical privacy concerns as they share sensitive patient data with cloud providers and third-party models. This study presents HealthAgent, a privacy-preserving LLM agent framework that protects both user queries and multi-modal sensor data through homomorphic encryption. HealthAgent enables an LLM orchestrator to coordinate specialized AI models for complex health assessments while processing all data in encrypted form. The system achieves 95% task decomposition accuracy with 10s latency, demonstrating that strong privacy guarantees can be maintained without sacrificing real-time performance in health monitoring applications. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.

Large Language Models (LLMs) are increasingly used to generate dynamic dialogue for game NPCs. However, their integration raises new security concerns. In this study, we examine whether adversarial prompt injection can cause LLM-based NPCs to reveal hidden background secrets that are meant to remain undisclosed. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2026.

Federated learning has rapidly advanced as a privacy-preserving, distributed machine learning methodology. Protecting the intellectual property rights of federated models, however, poses significant challenges. Existing backdoor-based watermarking techniques suffer from complexity, inefficiency, and implementation challenges. In this paper, we present AggreMark, a novel watermarking approach that incorporates parallel server-side training and watermark embedding during the weighted averaging phase of federated learning. This method does not alter the original model structure or parameters related to the original learning task and adds no additional time overhead, ensuring its practicality for real-world applications. AggreMark is easy to implement and maintain, and it robustly preserves the model's high performance. We validate the method's effectiveness and robustness against pruning-based watermark removal attacks through experiments on CIFAR-10. © 2024 IEEE.