I rendered 1,418 confusables over 230 fonts. Most aren't confusable to the eye
9 hours ago
- #Typography
- #Unicode
- #Security
- The tool 'confusable-vision' was developed to measure visual similarity of Unicode confusable pairs by rendering and comparing them across 230 fonts.
- 96.5% of confusable pairs score low on visual similarity, but 82 pairs are pixel-identical in at least one font, posing a high risk.
- Cyrillic homoglyphs are particularly dangerous, being pixel-identical to Latin characters in many standard fonts, making them indistinguishable.
- The study uses SSIM (Structural Similarity Index Measure) for deterministic and reproducible glyph comparison without relying on machine learning.
- Font choice significantly impacts confusable risk, with some fonts like Phosphate and Copperplate having high danger rates due to geometric designs.
- The findings suggest that confusable detection systems should consider rendering context and font-specific risks rather than relying solely on Unicode data.
- The tool is open-source and reproducible, with detailed documentation and scored data available for further analysis.