IAAR-Shanghai/MemPrivacy-4B-RL
MemPrivacy-4B-RL is a lightweight, privacy-preserving model developed by IAAR-Shanghai from the Qwen3-4B base model, optimized through reinforcement learning. It excels at high-precision privacy information extraction, surpassing models like GPT-5.2, and classifies sensitive data using a four-level privacy taxonomy. Designed for edge-cloud agents, it replaces privacy-sensitive spans with semantically structured placeholders to preserve utility while ensuring data privacy. This model is ideal for personalized memory management in resource-constrained edge environments, offering efficient and adaptive privacy protection.
Loading preview...
MemPrivacy-4B-RL: Edge-Optimized Privacy Preservation
MemPrivacy-4B-RL, developed by IAAR-Shanghai, is a specialized 4-billion parameter model built upon Qwen3-4B and enhanced with reinforcement learning. Its core function is to provide privacy-preserving personalized memory management for edge-cloud agent systems.
Key Capabilities
- High-Precision Privacy Extraction: The model accurately identifies and extracts privacy-sensitive information from conversational text, outperforming larger general-purpose models like GPT-5.2 and Gemini-3.1-Pro in this specific task.
- Four-Level Privacy Taxonomy: It classifies identified privacy-relevant content into four distinct levels (PL1-PL4) based on identifiability, potential harm, and operational exploitability, allowing for fine-grained, user-configurable protection policies.
- Semantic Utility Preservation: Instead of aggressive masking, the model replaces sensitive data with semantically structured, type-aware placeholders (e.g.,
<Email_1>). This ensures that cloud agents can still understand the context and relationships within the data, crucial for effective memory formation and personalization. - Edge-Optimized Efficiency: Designed for deployment on resource-constrained edge devices, MemPrivacy-4B-RL maintains high accuracy while significantly reducing inference latency compared to massive general-purpose LLMs.
Good For
- Developers building edge-cloud agent systems requiring robust data privacy.
- Applications needing high-precision identification and classification of sensitive user data.
- Use cases where semantic integrity of data must be maintained even after privacy protection.
- Scenarios demanding efficient, low-latency privacy processing on local devices before cloud transmission.