VOGUES: Validation of Object Guise using Estimated Components


Raymond Muller, Purdue University; Yanmao Man and Ming Li, University of Arizona; Ryan Gerdes, Virginia Tech; Jonathan Petit, Qualcomm; Z. Berkay Celik, Purdue University


Object Detection (OD) and Object Tracking (OT) are an important part of autonomous systems (AS), enabling them to perceive and reason about their surroundings. While both OD and OT have been successfully attacked, defenses only exist for OD. In this paper, we introduce VOGUES, which combines perception algorithms in AS with logical reasoning about object components to model human perception. VOGUES leverages pose estimation algorithms to reconstruct the constituent components of objects within a scene, which are then mapped via bipartite matching against OD/OT predictions to detect OT attacks. VOGUES's component reconstruction process is designed such that attacks against OD/OT will not implicitly affect its performance. To prevent adaptive attackers from simultaneously evading OD/OT and component reconstruction, VOGUES integrates an LSTM validator to ensure that the component behavior of objects remains consistent over time. Evaluations in both the physical domain and digital domain yield an average attack detection rate of 96.78% and an FPR of 3.29%. Meanwhile, adaptive attacks against VOGUES require perturbations 30x stronger than previously established in OT attack works, significantly increasing the attack difficulty and reducing their practicality.

Open Access Media

USENIX is committed to Open Access to the research presented at our events. Papers and proceedings are freely available to everyone once the event begins. Any video, audio, and/or slides that are posted after the event are also free and open to everyone. Support USENIX and our commitment to Open Access.