Gmail adopts deep learning to scan and block malicious attachments

Thanks to its ubiquity via Google's other services, Gmail accounts have become prime targets for malicious or even illegal activities. It isn't just spam or phishing scams either, as emails can also include files that, when downloaded and opened, could compromise users' devices and files. That's why Google is ramping up its efforts to block these attachments from even reaching your Gmail inbox in the first place and it's doing it with a little help from Google's favorite machine learning special sauce.

It shouldn't come as a surprise that Google scans emails that pass through and especially those that land on its servers. That's part of what you agree to when using the service, whether for free or otherwise, both to improve Gmail as well as help filter out spam or even organize your email into categories. And it isn't just the email itself that's being processed but even attachments that may actually do more harm on their own.

Google boasts that it is able to process 300 billion attachments each week but the strategies that hackers and criminal elements employ continually evolve, sometimes faster than any static system could keep up with. It's cases like that when AI and machine learning are more suited since they can analyze and learn faster than humans, especially when large volumes of data are involved.

The company deployed a new deep learning document scanner at the end of 2019 and it is reporting a marked but not exactly a staggering increase in detection rate, about 10%. In the case of what it calls "adversarial, bursty attacks", the success rate has increased by a whopping 150%.

The new attachment scanner, however, is still at its infancy with plenty of room for growth. For one, it only focuses on documents since they compromise 58% of Gmail malware. Google is also so far only scanning Office documents, the most common carrier of such malware. The company hasn't mentioned anything regarding how it is protecting users' privacy with this deeper scanning method.