Google has announced a major upgrade to its spam filtering system in Gmail. This upgrade, called RETVec (Resilient & Efficient Text Vectorizer), is touted as the biggest upgrade in several years. RETVec utilizes a new text analysis system that can understand modified advertisements intended to evade detection. These modifications can include the use of special characters, emojis, misspellings, and other techniques that make the ads readable to humans but potentially confusing to computers.
For example, some ads may use words like “C0NGRATULATIONS” instead of “Congratulations” or “Jᴀ̲ᴄ̲ᴋ̲pot” with underlined characters that are not normally paired using Unicode. RETVec is a trained model that has been trained with modified text, including additions, deletions, misspellings, substitution of special characters for letters, and other techniques in UTF-8 characters across more than 100 languages. The results are not in a keyword lookup table, as the size would be significantly larger than the traditional method, requiring the conversion of characters with approximately 200,000 parameters. However, the open-source RETVec model is now available on GitHub.
TLDR: Google has upgraded its spam filtering system in Gmail using RETVec, a new text analysis system capable of detecting modified advertisements that try to evade detection by using techniques like special characters and misspellings. The RETVec model, trained with modified text in over 100 languages, is an open-source project available on GitHub.
Leave a Comment