Google has rolled out a new text classification system to Gmail, bringing one of the biggest security upgrades in recent years to the mail service.
It involves the classification system RETVec, short for Resilient & Efficient Text Vectorizer. A mail service like Gmail relies on classification models to identify malicious content. Think about phishing attacks, inappropriate content and other scams.
To strengthen the text classification system, Google developed RETVec. RETVec helps improve classification performance of models and reduce computational power costs. In the image below, Google shares figures on how RETVec improves spam filtering.
Google says it has been testing RETVec extensively internally for the past year to determine its usefulness for classification. The system proved to be very effective for security and anti-abuse purposes. In particular, RETVec performs well when it replaces an existing component in the Gmail spam classification function. Those performance improvements can be seen in the image using the figures provided.
“Additionally, using RETVec reduced the TPU usage of the model by 83%, making the RETVec deployment one of the largest defense upgrades in recent years,” Google said.
RETVec achieves the performance gains primarily by using a lightweight model of about 200,000 parameters. This allows Google to reduce the size of the Transformer model used, resulting in equal or better performance.