E2PRIV: Privacy-Preserving Event-to-Video Reconstruction with Face Anonymization

Event cameras are increasingly adopted in humancentered applications due to their ability to capture high-speed motions without blur and fine-grained details such as facial micro-expressions. While their unconventional pixel-based data structure is often considered privacy-preserving, event-to-video reconstruction algorithms challenge this assumption by revealing identifiable faces from event streams, commonly known as a visualization attack. Since these sensors offer high temporal and dynamic resolution, reconstructed faces often retain fine details, increasing the risk of sensitive information leakage. To address this challenge, anonymization techniques have been explored, but they are typically applied post hoc as independent processing steps. Accordingly, we propose E2PRIV, the first event-based reconstruction model that integrates anonymization directly within the reconstruction process. By incorporating dynamic Gaussian blurring and noise injection during training, our model effectively obscures identifiable facial features while balancing the trade-off between privacy and usability. We evaluate our anonymization approach through face detection and verification, demonstrating a significant reduction in both detected faces and identity matching rates. To further validate its effectiveness, we study the impact of anonymization on action recognition and prove that anonymized frames maintain performance without degradation, ensuring that they remain functional for human-centered downstream applications. Code available at: https://github.com/MyraAdra/e2priv.