Handwriting examination techniques have traditionally been developed under the assumption that all handwriting within a document comes from a single writer, which imposes fundamental limitations on detecting locally mixed forgeries within the document...
Handwriting examination techniques have traditionally been developed under the assumption that all handwriting within a document comes from a single writer, which imposes fundamental limitations on detecting locally mixed forgeries within the document. To overcome this, the present study proposes a Multiple Instance Learning (MIL)–based framework for the automatic detection of partial forgeries in a single document. The proposed framework combines deep learning–based feature extraction with metric learning to obtain fine-grained embeddings of writer-specific characteristics, and applies these embeddings within an MIL algorithm to determine whether a document has been forged. Experimental results show that the framework achieves meaningful performance not only in sentence-level forgery scenarios with rich information, but also in more challenging word-level local forgery scenarios. Detailed analyses reveal that detection performance improves as the proportion of forged instances increases, and additional experiments identify the data scale required for stable convergence of the model’s performance. Furthermore, by using an attention mechanism to visually highlight suspicious regions, the framework enhances interpretability beyond simple binary decisions.
In conclusion, this study demonstrates that the proposed MIL framework learns solely from bag-level document labels without explicit instance-level annotations indicating forged locations, while still detecting within-document writer mixtures effectively.