Neural Network for Adversarial Deep Learning in Wireless Systems
AssigneeNortheastern University
InventorsTommaso MELODIA, Francesco RESTUCCIA, Salvatore D'ORO
Life of this patent
Estimating the term…
Abstract
A method of determining a response of a radio frequency wireless communication system to an adversarial attack is provided. Adversarial signals from an adversarial node are transmitted to confuse a target neural network of the communication system. An accuracy of classification of the incoming signals by the target neural network is determined.
Claims21
A method of determining a response of a target wireless communication system to an adversarial attack, comprising:
- (a) providing, at an adversarial node, an adversarial processing unit operative to generate adversarial signals to confuse the target wireless communication system, wherein:
- the target wireless communication system includes a target neural network operative to classify incoming waveforms of incoming signals received from a plurality of transmitters, and
- the adversarial signals are based on an attack strategy comprising generating the adversarial signals with waveforms that cause the target neural network to misclassify legitimate signals;
- (b) transmitting the legitimate signals from one or more legitimate transmission nodes to the target wireless communication system;
- (c) transmitting the adversarial signals from the adversarial node to the target wireless communication system during transmission of the legitimate signals;
- and
- (d) determining an accuracy of classification of the legitimate signals by the target neural network.
A neural network training system for use with a target wireless communication system, the target wireless communication system including a target neural network operative to classify incoming waveforms of incoming signals received from a plurality of transmitters, the training system comprising:
- a radio frequency (RF) receiver operative to query waveforms from a target network and a transmitter operative to transmit signals to a receiver of the target network;
- an adversarial processing unit comprising:
- a trained discriminator operative to provide outputs comprising waveforms imitating legitimate transmitter devices of the target network, and
- a generator comprising a deep learning neural network including a plurality of neural network layers, the neural network layers including a plurality of finite impulse response layers, the generator operative to receive training inputs comprising the outputs from the trained discriminator and selected baseband waveforms, and to produce a training output comprising waveforms for a target class of waveforms of the target network,
- wherein the adversarial processing unit is further operative to transmit the waveforms of the training output to the receiver of the target network.
A method of training an adversarial neural network for use with a target wireless communication system, the target wireless communication system including a target neural network operative to classify incoming waveforms of incoming signals received from a plurality of transmitters comprising:
- (a) providing an adversarial processing unit comprising:
- a trained discriminator operative to provide outputs comprising waveforms imitating legitimate transmitter devices of the target network, and
- a generator comprising a deep learning neural network including a plurality of neural network layers, the neural network layers including a plurality of finite impulse response layers, the generator operative to receive training inputs comprising the outputs from the trained discriminator and selected baseband waveforms, and to produce a training output comprising waveforms for a target class of waveforms of the target network;
- (b) generating a waveform training batch based on a protocol of the target network;
- (c) assigning an embedded label to each waveform in the training batch;
- (c) inputting the training batch to the deep learning neural network;
- (e) outputting a training output from the deep learning neural network;
- and
- (f) transmitting waveforms in the training output to the target network.
Dependencies are inferred from the claim text.