Private Eye: On the Limits of Textual Screen Peeking via Eyeglass Reflections in Video Conferencing

### Overview With the widespread adoption of video conferencing during and after the COVID-19 pandemic, eyeglass reflections pose an unintended side-channel vulnerability. The presentation introduces "Private Eye," a project that characterizes and quantifies the limits of information leakage from eyeglass reflections captured by participants' webcams. Termed a "webcam peeking attack," this threat involves a remote conference participant eavesdropping on the screen contents of another attendee without controlling the victim's hardware.

### Key Topics - **Hardware & Computational Evolution:** Video cameras have transitioned from standard 720p laptop webcams and 1080p external cameras toward 4K streaming sensors. In parallel, adversaries can employ multi-frame super-resolution and image registration over consecutive frames to enhance reflection quality. - **Environmental & Optical Factors:** Information leakage is influenced by reflection pixel size (governed by screen-to-glass distance, camera resolution, glass curvature) and signal-to-noise ratio (affected by ambient room lighting). Notably, auto-exposure controls cause lighting intensity to impact reflection clarity non-monotonically, and stronger prescription glasses with flatter surfaces leak more detail. - **Text vs. Website Layout Recognition:** While reading small body text is difficult with existing hardware, identifying large texts or entire website layouts presents an immediate risk.

### Technical Details - **Evaluation Methodology:** The study combined controlled laboratory experiments (single-letter tests) with user studies involving 20 participants viewing HTML text of varied sizes. Amazon Mechanical Turk workers were used as a baseline to humanly recognize reflected text, complemented by multi-frame super-resolution. - **Text Leakage Limits:** With current webcams (720p/1080p), adversaries can read large fonts (roughly 56 pt or 10 mm physical height on screen, found on sites like Twitch). Predictive mathematical models indicate that future 4K webcams will allow peeking at most HTML header text, though body text remains challenging. - **Website Identification:** Attackers can identify website layouts and shapes rather than specific text. On the Alexa Top 100 dataset, the researchers achieved over 90% accuracy in identifying the visited website (such as Bing) using present-day webcams.

### Mitigations & Takeaways - **Individual Reflection Testing:** Platforms like Zoom could implement a per-session calibration test displaying known patterns to quantify current reflection risks and alert users. - **Software Filtering:** The authors developed a Python prototype that detects the eyeglass area in the camera feed and applies an adjustable blur filter before streaming. - **Principle of Least Pixels:** Treating this as an "over-sensing" issue where sensors capture more data than necessary, video conferencing systems should adopt a "principle of least pixels"—defaulting to lower resolutions and requiring deliberate user authorization for higher-capability video streams.