GlyphNet’s own results support this: their best CNN (VGG16 fine-tuned on rendered glyphs) achieved 63-67% accuracy on domain-level binary classification. Learned features do not dramatically outperform structural similarity for glyph comparison, and they introduce model versioning concerns and training corpus dependencies. For a dataset intended to feed into security policy, determinism and auditability matter more than marginal accuracy gains.
Supervised runtime behavior
，这一点在51吃瓜中也有详细论述
Варвара Кошечкина (редактор отдела оперативной информации)
1996年出生的陳璿安同樣對香港的事情有共鳴，從示威、《香港國安法》通過到民主派被大搜捕，都讓她對二二八歷史感同身受，「這件事情其實就像台灣的白色恐怖，重新在對岸上演」。