The open and vulnerable nature of wireless channels exacerbates security risks in integrated sensing and communication (ISAC) systems, especially when the sensing targets act as potential eavesdroppers (Eves), and these risks intensify with collusion among Eves. To address this challenge, this paper investigates a novel strategy for a robust reconfigurable intelligent surfaces (RIS)-assisted secure ISAC system, where an ISAC base station facilitates simultaneous secure communication with legitimate users and sensing of multiple targets that may serve as Eves. We examine two different interaction mechanisms among Eves, namely, non-colluding Eves (NCE) and colluding Eves (CE), under both perfect and imperfect channel state information (CSI) assumptions. For both mechanisms, we formulate the optimization problem of maximizing users’ sum secrecy rate by jointly designing the transmit beamforming and RIS phase-shifts. This optimization is subject to constraints on transmit power, sensing requirements, and unit-modulus RIS phase shifts. The resulting non-convex problems are solved via alternating optimization (AO) algorithms. Specifically, in order to handle the severely non-convex and coupled objective function and multi-link accumulated channel error constraints caused by CE as well as imperfect CSI, we employ the majorizationminimization algorithm and the S-procedure to convert these problems into tractable forms. Simulation results validate the effectiveness of our proposed algorithms. We highlight that, at the expense of a 15% reduction in the users’ sum rate, our proposed algorithm achieves up to a 185% increase in the sum secrecy rate. Furthermore, we quantify the sensing-security trade-off by analyzing the reduction of the sum secrecy rate induced by sensing requirements, and we reveal the impacts of various factors on the sum secrecy rate, such as RIS element number, channel estimation errors, and sensing thresholds.

Abstract Intelligent reflective surface (IRS) provides an effective solution for reconfiguring air‐to‐ground wireless channels, and intelligent agents based on reinforcement learning can dynamically adjust the reflection coefficient of IRS to adapt to changing channels. However, most exiting IRS configuration schemes based on reinforcement learning require long training time and are difficult to be industrially deployed. This paper, proposes a model‐free IRS control scheme based on reinforcement learning and adopts transfer learning to accelerate the training process. A knowledge base of the source tasks has been constructed for transfer learning, allowing accumulation of experience from different source tasks. To mitigate potential negative effects of transfer learning, quantitative analysis of task similarity through unmanned aerial vehicle (UAV) flight path is conducted. After identifying the most similar source task to the target task, parameters of the source task model are used as the initial values for the target task model to accelerate the convergence process of reinforcement learning. Simulation results demonstrate that the proposed method can increase the convergence speed of the traditional DDQN algorithm by up to 60%.