Google expanded its use of machine learning inside Chrome in 2022, applying local models to phishing detection, notification permissions and the browser interface.
The company reported that an on-device model had made Chrome 2.5 times better at identifying potentially malicious sites and phishing attacks since March. Local processing allowed the browser to react to current page signals without sending every interaction to a server.
The changes showed how small, focused models could improve an everyday product quietly.
Phishing Required Fast Detection
Phishing pages imitated trusted services to steal passwords, payment details or other information. Attackers changed domains and page designs quickly, limiting the value of static blocklists.
Machine learning could combine visual, textual and behavioural signals to identify suspicious pages that had not yet been widely reported.
Speed mattered because a warning arriving after credentials were submitted offered little protection. Running a model in the browser reduced network delay and worked even when cloud contact was limited.
Local Processing Supported Privacy
On-device analysis kept more browsing signals on the user’s machine. That was important because a complete browsing history could reveal health, politics, finances and personal relationships.
Local operation did not automatically guarantee privacy. Models could still produce telemetry, and browser settings determined what was shared with Google.
Clear documentation and controls were needed so users understood the difference between local detection and cloud-based Safe Browsing services.
Models Added a Layer, Not a Perfect Shield
No detector identified every malicious page. False negatives left users exposed, while false positives could block legitimate small businesses or new websites.
Chrome needed an appeals process and multiple signals before applying the most disruptive response. Warnings should explain the risk without training users to click through automatically.
Security also depended on updates. Attackers adapted their pages once they understood which patterns triggered detection.
Notification Prompts Became More Selective
Websites frequently asked permission to send notifications before a visitor understood their value. Many prompts were ignored, denied or abused for spam.
Chrome used machine learning to predict when a user was unlikely to grant permission and reduce unnecessary interruption. The browser could make the request less prominent rather than presenting a disruptive pop-up.
This improved usability and reduced opportunities for deceptive sites to secure ongoing access.
Personalisation Stayed on the Device
The model could learn from a person’s own responses to permission requests. Someone who consistently rejected notifications could receive fewer prompts.
Personalisation was useful when it respected a user’s pattern without creating a central behavioural profile. Local inference offered that possibility.
Users still needed an easy way to reverse the browser’s prediction and grant access to a chosen site.
The Toolbar Adapted to Context
Google also explored a toolbar that surfaced actions based on current needs. Sharing might appear when a user often shared links, while voice search could become more prominent for someone who used it.
Adaptive interfaces could reduce effort but also make controls feel unstable. A button that moved unpredictably imposed a new cognitive cost.
Changes worked best when they were subtle, explainable and did not hide essential security settings.
Small Models Could Have Large Reach
Chrome’s enormous user base meant a modest improvement could prevent many attacks. On-device models also avoided the expense and latency of sending every page to cloud infrastructure.
The approach required careful optimisation for memory, battery and older hardware. Security could not be reserved for users with the newest devices.
Fallback rules remained necessary when a device could not run the model effectively.
Attackers Could Target the Model
Adversaries might test page variations until they found designs that escaped detection. They could also use compromised legitimate sites, which looked trustworthy to a classifier.
Defence therefore combined model predictions with reputation systems, sandboxing, password protection and rapid incident response.
Google needed to monitor for uneven performance across languages and regions, since attackers often targeted communities with fewer security resources.
Chrome Made ML an Invisible Safety Feature
The 2022 update was not a general assistant or dramatic generative product. It used machine learning for bounded decisions where speed and scale mattered.
On-device execution offered a useful balance among protection, privacy and responsiveness. Its limitations were manageable because the model supported a layered security system rather than acting alone.
The enduring lesson was that some of AI’s most valuable consumer applications were nearly invisible: fewer dangerous pages, fewer intrusive prompts and a browser that adapted without exporting every signal to the cloud.
