Abstract
Event cameras have emerged as a compelling alternative to conventional frame-based RGB cameras in applications such as autonomous driving, robotics, and surveillance because of their high temporal resolution, wide dynamic range, and low latency. Unlike frame-based cameras, event cameras’ asynchronous, event-based data format inherently preserves privacy by recording only pixel-level brightness changes instead of complete visual scenes. However, similar to conventional cameras, event cameras remain vulnerable to security threats such as unauthorized data distribution and malicious manipulation. The absence of effective watermarking techniques further amplifies these risks. In addition, advances in deep neural networks have enabled the reconstruction of high-quality images from event data, undermining the inherent privacy advantage provided by an event camera. To mitigate these threats, we introduce ADVISE, a novel watermarking scheme to protect the privacy, ownership, and authenticity of event data. ADVISE employs a jointly trained deep CNN-based encoder–decoder architecture, in which the encoder embeds imperceptible watermarks into event data to preserve privacy in reconstructed images without degrading performance on downstream tasks such as gesture recognition, eye tracking and person re-identification, while the decoder accurately extracts the embedded watermarks for authentication. We conducted extensive experimental evaluations across multiple event datasets and a user study to assess the proposed watermarking scheme in terms of privacy preservation capability, detectability, robustness against various degradations, perceptibility and its impact on downstream task performance. Our results show that the embedded watermark is reliably extractable, achieving Bit Error Rates (BER) as low as 1.86% while remaining visually imperceptible to human observers. The watermark introduces only about 0.5% reduction in downstream task accuracy, demonstrating strong utility preservation. At the same time, ADVISE enhances privacy by significantly degrading the quality of images reconstructed from watermarked event data, reducing how much scene content can be recovered even with state-of-the-art event-to-image models.
Raveen Wijewickrama
Postdoctoral Researcher
Postdoctoral researcher at UT San Antonio working on mobile, wearable, and urban sensing systems and their security and privacy.