Authors: Assistant Professor V.Priyanandhini, Assistant Professor Vandana M
Abstract: Distributed machine learning frameworks encounter important privacy risks owing to the presence of data and model parame-ters across untrusted networks and nodes who may have malicious intents. In this paper, we describe our privacy-preserving framework which combines label retention on the master node, feature obfuscation using Lagrange interpolation based encod-ing, and differential privacy noise injection. Our methodology deals with privacy risks arising out of honest but curious work-ers, external snooping attacks, and heterogeneity problems like stragglers in a single solution. Our experimental analysis con-ducted on MNIST, CIFAR-10, Fashion MNIST, and simulated health care databases indicates that our framework offers com-parable accuracy within 1.6 percent reduction against non-private baseline solutions, besides ensuring strong privacy. It com-pares favourably with federated learning along with differential privacy in terms of 2-3 percent gain in accuracy and outpaces federated learning solutions based on homomorphic encryption in terms of computation time. The framework offers stable results under up to 40 percent straggler worker conditions.
