Large model-oriented dynamic confrontation defense and traceable content generation method and system

By employing dynamic adversarial defense and traceable content generation methods, and utilizing hybrid network structures and attack knowledge graphs to optimize the security defense of large models, the static response dilemma and capability fragmentation are resolved. This achieves efficient and covert adversarial defense and watermarking technology, thereby enhancing the security defense capabilities of large models.

CN122310491APending Publication Date: 2026-06-30HANGZHOU ANQUAN DIGITAL INTELLIGENCE TECH CO LTD
0 Cites 0 Cited by

Patent Information

Application Number
CN202610411497.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-03-31
Publication Date
2026-06-30

AI Technical Summary

Technical Problem

Existing large-scale security defense solutions suffer from static response dilemmas, capability fragmentation, and conflicting objectives, making it difficult to dynamically respond to zero-day attacks and variant attacks. Furthermore, watermarking technology cannot simultaneously meet the requirements of concealment, robustness, and reversibility.

Method used

We adopt a dynamic adversarial defense and traceable content generation method for large models. We detect requests through a hybrid structure network, perform multimodal purification and causal intervention, embed implicit watermarks, and construct an attack knowledge graph for system optimization.

Benefits of technology

It achieves real-time perception and dynamic adjustment, significantly improving the success rate of defense, balancing high-precision detection with low latency, enhancing model robustness and watermark concealment, and resolving the multi-objective contradictions in existing technologies.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention relates to a large-model-oriented dynamic confrontation defense and traceable content generation method and system, and the method comprises the steps: receiving a multi-modal request, extracting basic features, carrying out the detection through a first detector, carrying out the semantic and structural analysis of the request based on an output result, recognizing an attack type and a risk level, and carrying out the detection through a second detector. Performing a multi-mode purification operation and / or a causal intervention operation on request content, or directly inputting the request into a core large model to generate a modulation signal in a content generation and decoding process, superposing the modulation signal to a hidden state of the model, and embedding a hidden watermark; performing compliance check on the generated content, and correlating to generate metadata; analyzing and constructing an attack knowledge graph, and directionally optimizing and updating the first detector or the core large model based on an analysis result; the system comprises a request access and rapid detection module, an analysis and dynamic purification module, a security generation and implicit watermark embedding module, an output post-processing and metadata association module and a security information collection and closed loop evolution module.
Need to check novelty before this filing date? Find Prior Art