Trojan horse detection method and device, electronic equipment, medium and program product
By acquiring the deserialized instruction text sequence of a deep learning model and constructing risk analysis prompts, and then interacting with a Trojan detection model, the problem of difficulty in identifying complex Trojan behaviors in existing technologies is solved. This achieves efficient and low-cost model detection, improving detection accuracy and real-time performance.
Patent Information
- Application Number
- CN202511145047.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-15
- Publication Date
- 2026-06-02
AI Technical Summary
Existing Trojan detection methods are ineffective at identifying malicious instructions in deep learning models, especially those Trojan behaviors that have undergone mutation or have complex structures. Furthermore, they are costly to maintain and cannot meet the financial industry's requirements for high reliability and accuracy in ensuring the security of the model supply chain.
By obtaining the deserialized instruction text sequence of the target model file, a risk analysis prompt word is constructed using an instruction analysis agent, and interacts with the Trojan detection model to generate risk judgment results, including risk level, category, and source. The pre-trained language model is fine-tuned using multi-source training corpus to improve detection accuracy and real-time performance.
It implements a lightweight model detection process, reduces operation and maintenance and resource costs, can identify complex Trojan behavior, improves the detection accuracy of unknown threats and variant attacks, and provides real-time and automated detection capabilities.