Machine learning–powered Android Trojans bypass script-based Ad Click detection

Researchers at Dr[.]Web have uncovered a new Android click-fraud trojan family that uses TensorFlow.js ML to visually detect and tap ads, bypassing traditional script-based methods. Distributed via Xiaomi’s GetApps, the malware operates in a hidden phantom WebView and loads models from remote servers, analysing screenshots to mimic real user behaviour and evade dynamic ads.

In phantom mode, Android.Phantom.2.origin uses a hidden WebView that, on command from the playstations[.]click server, loads a target site and a phantom JavaScript file containing ad-automation logic and TensorFlowJS, with a trained model downloaded remotely. Signalling mode employs WebRTC, with dllpgd[.]click acting as a signaling server to stream a live video of the virtual browser to attackers who can control it remotely.

The campaign also tracks infections across multiple games believed to be from a single developer, SHENZHEN RUIREN NETWORK CO., LTD., and updates in October added Android[.]Phantom.5, a dropper carrying Android.Phantom.4.origin.

Beyond Xiaomi, the malware spreads via Moddroid and Apkmody, and through modified Spotify, YouTube, Deezer, and Netflix apps, with a Discord server hosting about 24,000 users promoting an infected Spotify X app; according to Dr[.]Web, these trojans can turn devices into bots for DDoS and other illicit activities. 22 January 2026