Google Tackles Gmail Spam with Tensorflow

Google reports Gmail is blocking 100 million extra spam emails per day following the implementation of Tensorflow, its open source, machine-learning framework, to supplement existing spam detection.

Machine learning isn't new to Gmail: Google has long been using machine-learning models and rule-based filters to detect spam, and its current protections have reportedly prevented more than 99.9% of spam, phishing, and malware from landing in Gmail inboxes. Today's attackers seek new ways to hit Gmail's 1.5 billion users and 5 million business clients with advanced threats.

Considering the size of Gmail's user base, 100 million extra messages doesn't seem like much. However, since it already blocks so much, the last remaining threats are toughest to identify.

Enter TensorFlow, an open source software library that developers can use to build artificial intelligence (AI) tools. It was developed by researchers and engineers from the Google Brain team within its AI division in 2015, and is used among companies including Google, Intel, SAP, Airbnb, and Qualcomm.

"We're now blocking spam categories that used to be very hard to detect," said Neil Kumaran, product manager for counter-abuse technology, in a blog post on the news.

TensorFlow protections complement Google's machine learning and rule-based protections to try and block the last 0.1% of spam emails from getting through. It supplements current detection by finding image-based messages, emails with hidden embedded content, and messages from newly created domains that may try to hide a low volume of spam emails within legitimate traffic.

Unlike rule-based spam filters, machine-learning models hunt for patterns in unwanted emails that people may not catch. Every email has thousands of defining signals, each of which can help determine whether it's legitimate. TensorFlow helps weed through the chaos and spot spammy emails that seem real, as well as emails that have spam-like qualities but are authentic.

Kumaran says TensorFlow also helps with personalizing spam protections for each user. The same email could be considered spam to one person but important information to another.

Applying machine learning at scale can be complex and time-consuming. Google is aiming to simplify the process with TensorFlow, which also adds the flexibility to train and experiment with different models at the same time in order to choose the most effective, instead of doing so one at a time.

Still, Gmail security will continue to pose a major challenge for Google. A new report shows how attackers are abusing "dots don't matter," a longstanding Gmail security feature, to create fraudulent accounts on websites and use variations of the same email address.

Confidential Computing: Google Buckles Down on Asylo
Google reports it's investing in confidential computing, which aims to secure applications and data in use, even from privileged access and cloud providers. In addition to today's Gmail news, Google has published an update on Asylo, an open source framework it introduced in May 2018 to simplify the process of creating and using enclaves on Google Cloud and other platforms.

The adoption of confidential computing has been slow going due to dependence on specific hardware, complexity around deployment, and lack of development tools to create and run applications in these environments. Asylo makes it easier to build applications that run in trusted execution environments (TEEs) with different platforms – for example, Intel SGX.

Google anticipates in the future Aslo will be integrated into developer pipelines, and users will able to launch Asylo apps directly from commercial marketplaces. However, confidential computing is still an emerging technology and enclaves lack established design practices.

To accelerate its use, Google is starting a Confidential Computing Challenge, a contest in which developers can create new use cases. Applicants have until April 1 to submit essays describing a novel use case for the tech.

Related Content: