New Patent Judges Apps by the Company They Keep to Catch Malware

By Patentlyze Team · Updated Jul 10, 2026

Most malware detectors look at an app in isolation. Microsoft's new patent flips that around, teaching a model to judge whether an app is dangerous by looking at which other apps share its DNA.

How Microsoft's guilt-by-association malware detector works

Imagine a neighborhood watch that flags a new resident not because of anything they've done, but because three of their closest friends have criminal records. That's roughly the logic behind this patent.

Microsoft describes a system that takes a large pool of apps, some known to be safe and some known to be malicious, and maps out which ones are "neighbors" of each other based on shared characteristics. Maybe they use the same code libraries, request the same device permissions, or connect to the same network addresses. The system then trains an AI model on those neighborhood relationships, not just on each app alone.

The key output is a reputation score for specific app traits, like a particular permission or a shared developer certificate. If that trait keeps showing up in apps that are malicious, it earns a bad reputation, and any new app carrying it gets flagged faster. It's a way of making the whole catalog of known apps teach the system something useful about every new one.

How the model scores features across neighboring app clusters

The patent describes a pipeline with several distinct steps:

• Feature vectors: For each app, the system builds a list of measurable traits (called a feature vector). These might include permissions requested, URLs contacted, code signatures, or developer identifiers.
• Neighbor mapping: The system identifies "neighboring" apps for each app in the dataset, essentially apps that resemble it across those traits. Think of it as clustering similar apps together.
• Graph-aware model training: A threat detection model is trained not just on an individual app's features but on the features of its neighbors too. This is similar to graph neural network approaches, where context from connected nodes shapes predictions about each node.
• Feature weight extraction: After training, the model outputs feature weights, numerical values reflecting how much each trait contributed to a threat classification. A weight close to zero means that trait didn't matter much; a high weight means it was important.
• Reputation scoring: For each specific feature value (say, a particular API call or certificate), the system multiplies the app's own reputation score by the feature weight, then aggregates those products across all apps where that value appears. The result is a reputation score for the entity itself, independent of any single app.

That final reputation score means future apps can be evaluated partly by whether they share traits with historically bad actors, even before behavioral analysis runs.

What this means for Windows and enterprise security tools

Windows Defender and Microsoft's enterprise security platform Defender for Endpoint handle threat detection at enormous scale, processing millions of app evaluations daily. A system that can pre-score individual app traits, like a suspicious certificate or a rarely-used permission, means the model can make faster, more confident calls on apps it has never seen before.

For your device, this could mean malware gets caught earlier in the review pipeline, before it ever reaches a download page or slips through an enterprise app store. The broader signal here is that Microsoft is investing in relational threat intelligence, where the full catalog of known-bad software actively informs judgments about unknown software, rather than each app being evaluated in a vacuum.

Editorial commentary on a publicly published patent application. Not legal advice.