Secure Peer to Peer Learning Using Auto Encoders
Anirudh Kasturi, Akshat Agrawal, Chittaranjan Hota · 2022 IEEE International Symposium on Smart Electronic Systems (iSES) · 2022
We need machine learning algorithms that can use clients' data to create tailored models while adhering to tight privacy rules because people are concerned about privacy as more and more personal devices are being connected to the internet. Here, we present a robust method with a provable convergence rate to solve the issue, as mentioned earlier, in a distributed (peer-to-peer) asynchronous setting. This paper introduces a generic framework that trains an autoencoder on each client and then transmits the encoded data to the remaining clients. In a single round of communication, each client compiles data from several clients and trains a deep-learning model on the encoded data. In this work, we detail the architecture and implementation of the proposed approach and assess its efficacy in light of its comparison to the best-in-class central training and federated learning (FL) algorithms. Compared to FL and central training algorithms, our experiments on various datasets reveal that our proposed technique yields comparable accuracy. Compared to FL, our results demonstrate that the suggested approach reduces the quantity of data transfer by over 95 %, saving a significant amount of bandwidth.