Video Chat as a Spy: A Gaze Driven Side-channel on Concurrent Apps in Mobile Multi-window Environments

  1. 1Shandong University

IEEE Transactions on Dependable and Secure Computing 2026

01 Motivation & Threat

image.png

Video calls may keep the front camera active while users interact with another app.

A remote participant can exploit subtle eye movements to infer PINs and ongoing activities.


Only the peer-side video stream is needed.


No privileged permissions or local malware.

No user/device calibration or device-class knowledge.

02 How XGazeLeak Works

Estimate relative gaze trajectories from video, then use two inference branches:

fixation clustering and keypad matching for PINs; temporal modeling for activities.

Five activity classes: video watching, map browsing, text reading, text-image reading, and typing.

03 Key Findings

25 participants per task; primary devices: Huawei P40 Pro and iPad Pro.

47.16%

Mean Top-5 PIN accuracy

across PIN lengths

89.27%

Mean five-class activity accuracy

with 30-second windows

Your gaze can reveal what happens beyond the video call.

Original manuscript figures: iPad PIN results (left); video-chat activity results (right).