Face anonymization in driver monitoring systems

The use of dash-cams in fleet management provides valuable insights into driver behavior but raises significant privacy concerns regarding compliance with several legal frameworks across the world such as the GDPR and the CCPA. While removing personal data is legally required, standard anonymization techniques such as Gaussian blurring often degrade the image quality to a point where it becomes useless for machine learning tasks. This work explores the effects of a state-of-the-art face anonymization technique on driver images provided by Verizon Connect. We evaluate both the anonymization quality and the preservation of critical features like the subject's head pose and gaze direction. Additionally, we investigate the impact of training machine learning models on anonymized data, using Verizon Connect's existing production model as a baseline. Lastly, we evaluate the temporal identity consistency when this method is applied to every frame of a video. These experiments were executed using the anonymization technique proposed in the Face Anonymization Made Simple paper. For each experiment we tested two paradigms: random identity generation (standard anonymization) and targeted identity replacement (face-swapping). Our findings show that the face-swap approach achieves better results in terms of the effectiveness of the anonymization and the identity consistency in video data while obtaining comparable results across all other metrics. Additionally, classification models trained on anonymized data showed no significant degradation in performance, suggesting that these techniques are a viable solution to de-identify subjects while retaining the data utility for machine learning tasks.