DiffSketch: Using Count Sketch to give distributed learning both communication compression and differential privacy

DiffSketch uses Count Sketch to compress messages transmitted in distributed learning, simultaneously achieving communication efficiency and provable differential privacy. The authors prove that Count Sketch itself has the differential privacy property, and evaluations show that the method can provide a strong privacy guarantee of ε=1, reduce communication volume by 20-50 times, with only a small decrease in accuracy. Under the same privacy guarantee and compression, its absolute test accuracy is 5%-50% higher than baselines that handle privacy and communication separately.