Gmail wants to win the war on junk mail, adds AI spam detection

The big picture: Gmail’s latest update introduces AI-powered spam detection to protect users from unsolicited junk mail. This upgrade not only enhances user security, but also significantly improves the spam detection rate according to the search giant.

The world’s most popular email service has rolled out what might be its most significant update in years. Aimed at outwitting spammers through advanced AI, Gmail now features a new text classification system called RETVec (Resilient & Efficient Text Vectorizer).

This new spam detection technology is a step forward in the fight against adversarial text manipulations, which is a common tactic used by spammers to bypass traditional filters. Spammers have long exploited the limitations of filters by using special characters, emojis, and other deceptive techniques to make their emails appear legitimate while evading detection.

Gmail’s new spam filter tech put to the test

These manipulations often include homoglyphs, which are characters resembling standard alphabet letters. But these are, in fact, a lot different, which makes it challenging for conventional filters to recognize them as spam. Developed by Google, RETVec addresses this challenge head-on. It’s a machine-learning model that uses visual similarity rather than relying solely on character recognition.

The new approach allows Gmail to understand and filter out emails that employ deceptive tactics, significantly improving its ability to block spam and phishing attempts. Google has been pioneering AI in email for some time, which was kicked off by an AI feature that helped write emails all the way back in 2018.

According to Google, introducing RETVec has improved spam detection rates by 38% and dropped the false positives by nearly 20%. So, users are less likely to see spam in their inboxes, making it less likely they’ll miss important, legitimate emails that are mistakenly marked as spam.

RETVec’s lightweight word embedding model, made up of 200,000 parameters, lets you reduce the size of the Transformer model without compromising performance. This efficiency is crucial for applications on mobile and web-based devices.

Gmail’s spam filtering efforts goes back years, as it began blocking an extra 100 million daily spam emails in 2019 using TensorFlow. With RETVec, Google hopes to set a new benchmark in email security.