Preprint reports smaller vehicle learning uploads and faster routing in simulation

A preprint reports that one-bit model uploads reduced authenticated uplink size by 92.4% versus full-precision updates in simulated cooperative vehicle routing. Mean verified updates aggregated per round increased from 19.6 to 26.5. The researchers paired vehicles with complementary shared randomness to reduce error when combining private model updates. Their evaluation coupled traffic simulation with packet-level simulation of communication authenticated using post-quantum methods. At a parameter-level privacy budget of 1, the method reduced mean travel time by 14.3% versus independent quantization at the same budget and communication cost. Privacy results showed a tradeoff: at a budget of 0.3, a vehicle-origin inference attack achieved 30% accuracy with correlated quantization versus 22% with independent quantization after 100 rounds.