The remarkable prompting capability of large language models (LLMs) offers substantial convenience to users across diverse backgrounds. Nevertheless, as the sensitive information within prompts is inevitably exposed to LLMs, caution must be exercised to preserve privacy. Among various studies, text anonymization is considered an effective approach to preventing privacy leakage in prompts through text substitution. However, existing works overemphasize privacy while overlooks preserving contextual integrity, degrading semantic consistency. To address these concerns, this paper introduces a context-sensitive prompt privacy-preserving framework, namely Adaptive Linguistic Sanitization and Anonymization (ALSA). In specific, ALSA incorporates a three-dimensional scoring mechanism to dynamically quantify the substitutability of each word within a prompt by integrating the Privacy Leakage Risk Score (PLRS), the Contextual Information Importance Score (CIIS), and the Task Relevance Score (TRS). Subsequently, a clustering technique is adopted to dynamically determine the threshold for assigning an anonymization action (i.e., Retain, Replace, Encrypt, or Delete) by balancing privacy, semantics, and task relevance. Extensive experiments on five benchmark datasets validate the superiority of ALSA over state-of-the-art baselines in terms of accuracy, privacy preservation, and semantic integrity.

Abstract Traditional maritime search and rescue methods mainly rely on manual operation, which takes a long time to identify and results in poor search and rescue results. This paper applies computer vision technology to the field of maritime distress object detection and proposes an improved object detection algorithm YOLOv5s‐EFOE based on the YOLOv5s algorithm. Firstly, the authors change the detection head of the YOLOv5s algorithm to use a mixed channel strategy to build a more efficient decoupling head, reducing the number of 3 × 3 convolutional layers in the middle of the decoupling head to only one. The width of the head is scaled by the width multipliers of Backbone and Neck. Secondly, in conjunction with the SimOTA matching algorithm, the positive samples of the target to be tested are dynamically allocated to improve the recognition ability of different targets. Finally, considering the low pixel value of the maritime distress target from the perspective of unmanned aerial vehicles (UAV), the loss function CIoU in the original YOLOv5s is changed to EIoU. EIoU not only considers the distance and aspect ratio of the centre point, but also considers the true difference between the width and height of the prediction box and the real box, which improves the prediction accuracy of the anchor box. Experiments are conducted on a subset of the public dataset SeaDronesSee. The of the YOLOv5s‐EFOE model proposed in this paper reached 75.9%, reached 79.9%, and reached 44.5%. These indicators are superior to the original YOLOv5s model, YOLOv7 series models, and YOLOv8 series models. Compared with the original model, the YOLOv5s‐EFOE model has increased the by 5.4%, by 5.6%, and by 4.6%, improving the difficult to detect and missed detection situations. This model can be deployed on UAVs and can effectively identify maritime distress target, achieving the purpose of search and rescue.