Google Implements On-Device AI to Combat Scams on Chrome and Android

Google has initiated the integration of Gemini Nano, its on-device large language model (LLM), within the latest version of Chrome to enhance defenses against online spam, scams, and phishing attacks.

With the launch of Chrome version 137, Google revealed its testing of the on-device LLM as an extra security layer within the Enhanced Protection mode of the browser’s Safe Browsing functionality. Chrome’s Enhanced Protection mode represents the most stringent level of Google Safe Browsing, purportedly offering double the protection against phishing and scams compared to its Standard Protection mode.

At present, the application of Gemini Nano is focused on countering tech support scams specifically on desktop versions of Chrome. Looking ahead, Google aims to broaden its capabilities to identify various types of scams, including those related to package tracking and unpaid tolls. Moreover, plans are in place to introduce this feature to Chrome on Android devices later in 2025.

Gemini Nano's Role in Mitigating Tech Support Scams

The integration of Gemini Nano with Chrome’s Safe Browsing service utilizes the LLM to generate signals that enhance the reliability of determinations made about potentially harmful sites, such as those involved in tech support scams. When users navigate to a site that may be malicious, the on-device Gemini Nano LLM analyzes the content to extract security signals, such as intent. These insights are relayed to Safe Browsing, which ultimately decides whether the site represents a threat. If identified as dangerous, Chrome will issue a warning.

This methodology is crafted to uphold both performance and user privacy. The LLM is activated selectively and operates locally on devices, ensuring prudent management of system resources.

Only users who have enabled Enhanced Protection mode will have their LLM-generated security signals transmitted to Safe Browsing. In contrast, users on Standard Protection mode will receive indirect benefits via updates to blocklists.

Jasika Bawa, Group Product Manager for Chrome at Google, along with Phiroze Parakh, Senior Director for Engineering at Google Search, noted: “Gemini Nano’s LLM is ideally suited for this application due to its capacity to distill the diverse and complex nature of websites, which helps us rapidly adapt to evolving scam strategies.”

Advantages of On-Device LLM Implementation

Utilizing on-device LLMs yields several advantages beyond the reduced demands of remote hosting. Notably, this local execution empowers Google to identify threats as users encounter them in real time.

According to the Google Chrome Security team, the average lifespan of a malicious site is less than ten minutes, making on-device protection essential for detecting attacks that haven’t been crawled before. The local LLM model enables enhanced visibility by allowing Google to assess threats from users’ perspectives—considering that websites may display differently for different users, either for legitimate reasons like device compatibility and content personalization or for nefarious reasons such as evading detection by security mechanisms.

In conjunction with this development, Google has also introduced AI-powered notifications for Chrome on Android. If the on-device model identifies a concerning notification, users will receive a warning that gives them the option to unsubscribe or view the content that was blocked, as well as a choice to enable future notifications if they believe the alert was a mistake.

These enhancements come on the heels of Google’s previous introduction of AI-powered scam detection in the Android Messages app and the rollout of scam call flagging features the prior year.

Google’s Commitment to Combatting Online Scams

On May 8, Google unveiled its latest report, “Fighting Scams in Search,” asserting that investments in AI-enabled scam detection and classifier improvements have resulted in identifying 20 times more scam pages than previously achieved. The report highlighted an alarming trend of scammers impersonating airline customer service agents, with Google successfully reducing such scams in search results by over 80%.

This innovative approach underscores Google’s commitment to advancing cybersecurity measures through advanced artificial intelligence capabilities, enhancing user safety across its platforms and services.