Secure Collaborative Deep Learning Against GAN Attacks in the Internet of Things
Zhenzhu Chen, Anmin Fu, Yinghui Zhang, Zhe Liu, Fanjian Zeng, Robert H. Deng
Abstract
Deep learning makes the Internet-of-Things (IoT) devices more attractive, and in turn, IoT facilitates the resolution of the contradiction between data collection and privacy concerns. IoT devices with small-scale computing power can contribute to model training without sharing data in collaborative learning. However, collaborative learning is susceptible to generative adversarial network (GAN) attack, where an adversary can pretend to be a participant engaging in the model training and learn other participants' data. In this article, we propose a secure collaborative deep learning model which resists GAN attacks. We isolate the participants from the model parameters, and realize the local model training of participants via the interaction mode, ensuring that neither the participants nor the server would have access to each other's data. In particular, we target convolutional neural networks, the most popular network, design specific algorithms for various functionalities in different layers of the network, making it suitable for deep learning environments. To our best knowledge, this is the first work designing specific protocol against GAN attacks in collaborative learning. The results of our experiments on two real data sets show that our protocol can achieve good accuracy, efficiency, and image processing adaptability.