A joint training method, medium and system of a safety brain-like terminal and a clone terminal
Patent Information
- Application Number
- CN202510419956.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-03
- Publication Date
- 2026-10-16
AI Technical Summary
目前行业内关于AI模型的应用是基于大基数场景建立的模型，在实际应用过程中通过海量家庭数据对模型进行不断训练，即单一模型家庭协同训练，大多是处理普遍性场景和数据的模型，不能根据家庭细微差异化提供精准的服务
[0003]To make the application services of the model more accurately meet the actual needs of families, this invention aims to provide a joint training method for a secure brain-like terminal and a clone terminal to address the differentiated application needs of families and provide them with more practical, effective, and precise services. The specific implementation is as follows: ● First, a dynamic model pool for each family is constructed. Family data is acquired to form neurons, thereby building sub-models for various dynamic scenarios. While acquiring family data, the sub-models can learn autonomously. From data analysis and extraction to model training and collision, the sub-models can continuously fission and generate more sub-models to further refine the dynamic model pool. ● Second, the dynamic model pool (i.e., sub-models, also clone terminals) composed of numerous family neurons forms the brain-like terminal model (i.e., the main model). This improves the main model's adaptability to application scenarios while making it more sophisticated within the domain. It can also continuously train, collide, and learn autonomously, thereby providing effective, efficient, and precise services. ● Furthermore, while the sub-model and main model are being applied, they can also be jointly trained using data resources acquired from real-world scenarios (such as home device data, environmental data, entertainment data, energy consumption data, safety data, and human behavior data). This provides intelligent services to families and users while also offering a home security mechanism. Additionally, this invention fully considers the personalized and differentiated characteristics of each family due to factors such as home environment, human behavior, and lifestyle habits. Using a single model cannot provide accurate and effective services to families and users, and in some scenarios, the same data may produce different results in different families. Therefore, this invention has the following beneficial effects: ● Through joint training methods, the model training scenarios become more diverse. The model transforms from single-model family collaborative training to secure brain-like family joint training. Through continuous training, comparison, and listing across multiple scenarios, data, and applications, the model scenarios become more diverse and refined. ● Improved model processing capabilities and accuracy. Based on the continuous learning and improvement of the clone terminal in various application scenarios, the brain-like terminal it constitutes will have more accurate analysis, judgment, and processing capabilities for event handling. For example, the same data may produce different processing results in different families due to factors such as family environment, human behavior, and living habits. ● Using the "encryption factor" technology, random value encryption algorithms are applied to the family dynamic model pool (i.e., sub-model) and the brain-like terminal model (i.e., main model). The sub-model serves as a personalized encryption factor, which is combined with the common encryption factor of the main model. This allows the model to be continuously refined based on data acquisition, while also providing a multi-layered and effective security mechanism to comprehensively protect the security of family data.