Privacy-Preserving Parking Lot Surveillance Utilizing Homomorphic Encryption

Privacy is an important aspect of smart parking services provided by open urban platforms, because these systems process personal data such as license plates. In this work, we present a privacy-preserving computer vision system for a specific use-case of detecting unpaid parked cars. The main contribution is a system-level demonstration of an existing homomorphic encryption protocol CKKS applied to the parking lot surveillance. The use of homomorphic encryption enables encrypted license plate data processing on an external open urban platform without disclosing the specific content of the plates to platform owners. System implementation includes a convolutional neural network that operates on encrypted data. We evaluate the system experimentally and show that it reaches 84,27% accuracy in license plate recognition and scales linearly with number of cars on a parking space and number of requests albeit with a slow processing time per request due to computational costs associated with homomorphic encryption.