(4) Deep Defense of Distributed Deep Learning (D4L)

This project aims to achieve trustworthy deep learning (TDL)-based distributed computation among wireless connected Naval devices (sensors, UAVs, mobile phones, ships, etc.) through the seamless integration of cryptography, a complimentary classifier, and communications (C4). Distributed Deep Learning (DDL) runs in a distributed Naval network setting and is vulnerable to attacks that have knowledge of DDL meta-parameters. A simple generative adversarial network (GAN)-based model can cause the DDL to leak label information to adversaries. Key-based gradient encryption method does not include a complete DDL network security protocol. Blockchain may be used to protect DL; however, this approach does not have learning-based input control and cannot resist adaptive spiral attacks in a complex DDL setting. These elements enable a new challenging attack that is unique to the DDL environment: the adaptive grey-box spiral (AGBS) attack. Our project goal is to develop and demonstrate C4, a powerful new method that seamlessly integrates applied Cryptography, distributed deep Computation, a Complementary Classifier with ensemble Bayesian learning, and a Naval Communication scheme with Blockchain protocol.