MASTER Assignment
Real-Time Privacy-Preserving Collision Detection for UAVs using PSI
Type : Master M-CS
Period: November 2025 - April, 2026
Student: Haterd, van de R. (Rik, Student M-CS)
Date Final project: April 2, 2026
Supervisors:
Abstract:
Drones are increasingly deployed by independent operators in shared airspace, where exchanging trajectories improves safety but can reveal sensitive operational information. This thesis investigates whether Private Set Intersection (PSI), based on Oblivious Key-Value Stores (OKVS) and Vector Oblivious Linear Evaluation (VOLE), can support real-time, privacy-preserving drone-to-drone collision detection.
The proposed system represents short-horizon predicted trajectories as sets of discretised space-time voxels within an explicit safety radius. Collision detection is reduced to set intersection over these voxels and implemented using a VOLE-OKVS-based PSI protocol. A full prototype integrates trajectory prediction, encoding, and collision detection using PSI.
The evaluation examines how trajectory encoding choices affect computation time, communication cost, and alert timing across several operational configurations. Measured per-tick computation times range from single-digit milliseconds in performance-oriented settings to tens of milliseconds in more conservative configurations. Under modelled link capacities, some configurations meet the end-to-end real-time constraint of 50~ms, while others exceed them due to larger encoded sets. Multi-neighbour feasibility is driven primarily by concurrent scheduling rather than per-pair PSI cost.
We also investigate an updatable VOLE-OKVS PSI variant designed to reduce recomputation between time steps by patching the OKVS table. Although functionally correct, the updatable VOLE-OKVS PSI variant did not improve efficiency in practice, as small logical trajectory changes caused near-global table updates due to coupled OKVS entries. This removed the sparsity needed for efficient patching, pushed the cost close to a full rebuild, and introduced additional time-correlated leakage.
Overall, feasibility is driven primarily by trajectory representation, scheduling, and link capacity. In the evaluated host/SITL setup and under modelled compute and communication assumptions, VOLE–OKVS PSI meets the 50 ms per-tick requirement for some operating points. Deployment on embedded platforms and under realistic radio conditions remains unvalidated.