Gmail: new spam filter update featuring RETVec

Wow! This is pretty cool. Ron Amadeo from ARS Technica reports on a significant, AI-based spam filter update at Gmail. Gmail can now understand "adversarial text manipulations" using a new mechanism called RETVec (Resilient & Efficient Text Vectorizer), meaning that it basically renders graphics of everything written to compare it all for spam classification purposes, based on the words and other bits extracted from the message, basically regardless of how they're encoded. Using emojis to trick people in spam, because writing something it out in words will get you blocked? That might not fly now. Using cyrillic characters (that look visually similar to a user but look different to a text classifier) to try to skate by Gmail filters and hide the nature of what you're sending? Nuh-uh.

(Interestingly, I've known for a while now that Gmail can take issue with certain special characters or emojis in certain places in messages. I've not compiled some master list or anything, but I previously had a client whose mail was getting bounced intermittently at Gmail because of an emoji checkbox in the friendly from. It was a pain to troubleshoot! Nothing learned today makes me shy away from my prior guidance, which was and is: don't get silly with non-textual things.)

More from Google: Improving Text Classification Resilience and Efficiency with RETVec

[ H/T: Dan Deneweth ]

Post a Comment