Analysis and Mitigation of Local Feature Descriptor Reverse Engineering Attacks

Autonomous driving and augmented reality rely on localization based on user images, and the widely adopted local feature descriptors can, under certain conditions, be subjected to reverse engineering attacks that reconstruct images, posing a privacy risk. The paper uses a privacy threat model to characterize potential attackers, demonstrates such attacks on sparse feature maps under controlled conditions, and analyzes the vulnerability of FREAK, SIFT, and SOSNet. Finally, it evaluates mitigation techniques that select a subset of descriptors to balance privacy reconstruction risk and image matching accuracy, with results showing that similar accuracy can be achieved by disclosing less information.