Closed-Box Unlearning for Large Language Model-Enabled Internet of Everything

The Internet of Everything (IoE) represents a paradigm shift from device-centric connectivity to intelligence-centric ecosystem. To enable autonomous decision-making in dynamic environments, large language models (LLMs) have been integrated into IoE, leveraging the Model Context Protocol (MCP) to collect cross-device data for fine-tuning and transforming human natural language intents into structured tool calls that generate executable device instruction sequences. However, the massive data streams continuously generated in IoE environments often contain sensitive user privacy information, outdated low-quality data, and potentially malicious content, which can severely compromise the performance and trustworthiness of LLM-enabled IoE. Recently, LLM unlearning techniques have emerged as essential mechanisms for removing the influence of problematic data while preserving the model’s performance on retain data. However, existing unlearning approaches typically require full or partial access to the target LLM, which does not align with real-world IoE scenarios where only API access is available. To bridge this gap, we propose a closed-box unlearning framework for LLM-enabled IoE, without updating the target LLM parameters or accessing its internal logits. Our method employs a Retrieval-Augmented Generation (RAG)-enhanced dual-auxiliary architecture for forget-data detection, enabling dynamic selection of IoE response models to prevent leakage of forgotten data while maintaining optimal performance for legitimate queries. Comprehensive experimental evaluations on the TOFU dataset across multiple forgetting ratios demonstrate that our approach achieves superior unlearning effectiveness while preserving LLM utility.