Adversarial Deception on Deep-Learning Based Radio Waveforms Classification

Deep learning has achieved superior performance on radio signal modulation classification. However, its security and reliability are vulnerable due to its data-oriented learning strategy. In this paper, the vulnerability of deep neural network for radio classification is investigated. Based on the radar and communication waveforms, slight optimal perturbations are generated and added onto the original signals by searching for the saliency map. The simulated results show that the deep neural network will suffer degradation of classification accuracy due to adversarial deception both for radio signals with high and low signal-to-noise ratio.