Researcher(s)
- Ian Guy, Computer Science, University of Delaware
- Shukria Muhammadi, Cognitive Science, University of Delaware
Faculty Mentor(s)
- Chengmo Yang, Computer Engineering and Electrical Engineering, University of Delaware
- Jhon Abel Ordonez Ingali, Computer Engineering and Electrical Engineering, University of Delaware
Abstract
Virtual reality systems process highly sensitive user interactions, making privacy a critical concern. Existing security mechanisms primarily protect stored data and communication channels, yet physical side channels remain largely unexamined. During battery charging, a headset’s current consumption varies according to the executed workload, embedding information about the user’s actions in the power trace. An adversary capable of monitoring the charging current can exploit these variations to identify user activity, exposing a practical privacy threat that requires no access to device internals.
Prior work has demonstrated that such traces can distinguish application categories and individual websites, but these results depend on extensive hand-engineered preprocessing and manually designed feature extraction. This reliance constrains generalization across devices and collection conditions, and limits the practicality of the attack.
We address this limitation by learning discriminative representations directly from the measured signal. Current is acquired externally at the charging cable, with no software installed on the headset. Recordings were collected from six users under a common protocol, with data collection distributed across the team. Two complementary approaches are evaluated. For coarse-grained classification across application types, a one-dimensional convolutional network is trained on the raw current signal, with its architecture determined by automated neural architecture search rather than manual tuning. For fine-grained classification within a single category, transfer learning is applied to spectrogram representations of the same traces.
Across five application classes, the searched network classified unseen recordings with 97.5% accuracy while requiring approximately one quarter the model size of the transfer-learned baseline. For fine-grained discrimination among three shopping and search websites, transfer learning achieved 99.4% accuracy.
These findings indicate that publicly accessible charging infrastructure constitutes a viable surveillance channel for virtual reality users, and that deep learning substantially reduces the preprocessing required to exploit it.