Gmail's already robust spam filtering features are nearing perfection with machine learning
Initially developed for internal Google use, the TensorFlow open-source machine learning (ML) framework was released under the Apache 2.0 license more than three years ago. Since then, “teams and researchers all over the world” have produced a whopping 71,000 forks of the public code and other open-source contributions, establishing a strong community that makes it easy to quickly apply “new research and ideas.”
One of the more mainstream ideas was to further improve Gmail’s already impressive spam protection by adapting quickly to ever-changing, increasingly sophisticated techniques of spreading spammy messages.
Before implementing TensorFlow-powered protection features, Google’s effectiveness in blocking spam, phishing, and malware attempts was pretty great, but now it’s essentially flawless.
We’re talking a 99.9 percent success rate and an additional 100 million spam messages blocked every day with the help of machine learning technology. Google’s latest ML applications are purportedly capable of identifying “patterns in large data sets that humans who create the rules might not catch.”
Instead of focusing on just a few of an email’s characteristics that could seem “spammy” at a first glance or coincidentally fit with general spam-eliminating guidelines and red flags, ML provides a more complete view of suspicious messages, looking at all available signals before making a final determination. Thus, false positives are becoming rarer and rarer and spam categories that used to be particularly tricky to detect, like emails with hidden embedded content, are slowly vanishing.
The whole process is obviously automatic and blazing fast, although it's also continuously refined to enhance its speed and efficiency, working towards the utopian goal of ridding all 1.5 billion Gmail users worldwide of spam for all eternity.