In this paper, we propose a text anonymization method that utilizes a locally deployable small-scale language model (SLM) instead of a cloud-based large-scale language model (LLM) to address the privacy risks in user-generated content. We propose an anonymization framework based on self-reinforcement learning, called AgentStealth, which consists of an adversarial anonymization workflow utilizing in-context contrastive learning and adaptive utility-aware control, supervised learning of the SLM using high-quality data collected from the workflow, and online reinforcement learning utilizing the model’s internal adversarial feedback. Experimental results using two datasets show that the proposed method outperforms existing methods in terms of both anonymization effectiveness (+12.3%) and utility (+6.8%). Its lightweight design allows direct deployment on edge devices, thereby avoiding cloud dependency and communication-based privacy risks.
