A model structure similarity detection method and system based on strict matching of a Transformer block
By parsing, standardizing, merging LoRA layer shapes, and grouping matching methods, the problem of low detection accuracy in existing technologies is solved, achieving accurate similarity detection of deep learning models and protection of intellectual property rights.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- RUAN AN TECH CO LTD
- Filing Date
- 2026-04-08
- Publication Date
- 2026-07-17
AI Technical Summary
Existing model similarity detection methods ignore the interference of quantization layers when detecting the similarity of deep learning models, do not strictly group and match according to Transformer blocks, cannot accurately merge shapes, and do not consider the difference in the number of blocks, resulting in low detection accuracy and false positives.
By parsing and preprocessing the input model, standardizing layer names, merging LoRA adaptation layer shapes, grouping and extracting components by Transformer blocks, strictly matching each block level, applying a block number penalty factor, and calculating similarity scores, misjudgments are prevented.
It achieves accurate similarity detection of model structure, effectively distinguishes between normal similarity and plagiarism, adapts to different model variants, avoids local matching misjudgments, outputs bidirectional similarity, facilitates intellectual property auditing, and improves detection accuracy.
Smart Images
Figure CN122412979A_ABST