Systems similar to Gmail, YouTube and Google Play depend on textual content classification fashions to establish dangerous content material together with phishing assaults, inappropriate feedback, and scams. These sorts of texts are tougher for machine studying fashions to categorise as a result of unhealthy actors depend on adversarial textual content manipulations to actively try and evade the classifiers. For instance, they’ll use homoglyphs, invisible characters, and key phrase stuffing to bypass defenses.Â
To assist make textual content classifiers extra sturdy and environment friendly, we’ve developed a novel, multilingual textual content vectorizer known as RETVec (Resilient & Efficient Text Vectorizer) that helps fashions obtain state-of-the-art classification efficiency and drastically reduces computational price. Today, we’re sharing how RETVec has been used to assist shield Gmail inboxes.
Strengthening the Gmail Spam Classifier with RETVec
Figure 1. RETVec-based Gmail Spam filter enhancements.
Over the previous yr, we battle-tested RETVec extensively inside Google to judge its usefulness and located it to be extremely efficient for safety and anti-abuse purposes. In specific, changing the Gmail spam classifier’s earlier textual content vectorizer with RETVec allowed us to enhance the spam detection fee over the baseline by 38% and scale back the false constructive fee by 19.4%. Additionally, utilizing RETVec lowered the TPU utilization of the mannequin by 83%, making the RETVec deployment one of many largest protection upgrades in recent times. RETVec achieves these enhancements by sporting a really light-weight phrase embedding mannequin (~200k parameters), permitting us to scale back the Transformer mannequin’s dimension at equal or higher efficiency, and being able to separate the computation between the host and TPU in a community and reminiscence environment friendly method.
RETVec Benefits
RETVec achieves these enhancements by combining a novel, highly-compact character encoder, an augmentation-driven coaching regime, and the usage of metric studying. The structure particulars and benchmark evaluations can be found in our NeurIPS 2023 paper and we open-source RETVec on Github.
Due to its novel structure, RETVec works out-of-the-box on each language and all UTF-8 characters with out the necessity for textual content preprocessing, making it the best candidate for on-device, internet, and large-scale textual content classification deployments. Models skilled with RETVec exhibit sooner inference velocity on account of its compact illustration. Having smaller fashions reduces computational prices and reduces latency, which is essential for large-scale purposes and on-device fashions.
Figure 1. RETVec structure diagram.
Models skilled with RETVec will be seamlessly transformed to TFLite for cell and edge units, on account of a local implementation in TensorFlow Text. For internet software mannequin deployment, we offer a TensorflowJS layer implementation that’s obtainable on Github and you may try a demo internet web page operating a RETVec-based mannequin.
Figure 2. Typo resilience of textual content classification fashions skilled from scratch utilizing totally different vectorizers.
RETVec is a novel open-source textual content vectorizer that means that you can construct extra resilient and environment friendly server-side and on-device textual content classifiers. The Gmail spam filter makes use of it to assist shield Gmail inboxes towards malicious emails.
If you want to use RETVec to your personal use instances or analysis, we created a tutorial that will help you get began.
This analysis was performed by Elie Bursztein, Marina Zhang, Owen Vallis, Xinyu Jia, and Alexey Kurakin. We want to thank Gengxin Miao, Brunno Attorre, Venkat Sreepati, Lidor Avigad, Dan Givol, Rishabh Seth and Melvin Montenegro and all of the Googlers who contributed to the venture.