“Healthy surveillance” – KIT’s concept for privacy-preserving mask recognition AI

KIT proposes data protection-friendly monitoring in times of corona pandemic.

In response to the COVID-19 pandemic, many governments around the world recommended their citizens to wear masks as an effective countermeasure. Face masks are one of the best defenses against the spread of COVID-19, but their growing adoption brings to light another problem: monitoring of compliance. While manual inspections, e.g., at the entrance of restaurants or stores, are a possibility, these actions require manual labor, do not scale well and are difficult to enforce on larger spaces. In order to allow for an automated examination of the compliance, one could imagine to use surveillance solutions in combination with Artificial Intelligence (AI).

While this solution entails many upsides, e.g., scalability and automation capabilities, it needs to be in-line with the privacy regulations such as the GDPR and it needs to be understood by citizens to trust and accept the approach. The Karlsruhe Institute of Technology is now proposing an AI-based surveillance artifact which ensures both (a) privacy and (b) high performance of mask recognition. In a paper published in June, 2020, four KIT researchers  Niklas Kühl, Dominik Martin, Clemens Wolff, and Melanie Volkamer, show how a privacy-preserving mask recognition artifact could look like, and demonstrate different options for implementation and evaluate performances.

At the heart of their work is a conceptual deep-learning based Artificial Intelligence, able to achieve detection performances between 95% and 99% while maintaining privacy and protecting personal data. The paper clearly suggests that privacy-preserving mask recognition is well-feasible.

With their research, they contribute to the body of knowledge with three core aspects: First, they develop a novel artifact which can be utilized to allow for a privacy-preserving monitoring of mask coverage during a pandemic. Second, they evaluate different design choices on how to build the artifact and elaborate on their performances, strengths and weaknesses. Third and finally, they theorize on the trade-off between privacy preservation and AI performance—as AI performance decreases with increased privacy preservation and vice-versa.

Full article is available here.