A Review on Privacy Enhanced Distributed ML Against Poisoning Attacks

There is a higher demand for machine learning applications nowadays. To increase the accuracy of results, more and more training data is essential. To better increase the size of training data, it is necessary to distribute the machine learning workload to different nodes. Such learning systems are referred to as distributed learning systems. With the advancement of machine learning, processing all the data on a single server has become challenging. Therefore, machine learning tasks must be distributed across several servers, by converting centralized machine learning into distributed machine learning. Sometimes, attackers add impurities in the training phase to alter the trained ML model and it affects security of the model. Such attacks are called poisoning attacks. So, mitigating such attacks is still a challenging task. One of the most critical challenges is data and model security. This is a review paper focused on various security mechanisms for enhancing security in distributed ML systems. Privacy-preserving distributed learning mainly focuses on some basic cryptographic techniques. Among these methods, the hybrid method of encryption is used as the most suitable method. Due to complexity issue, it is better to integrate with particle swarm optimization techniques which will enhance the security as well as privacy of the ML systems.