POSITION-INVARIANT ADVERSARIAL ATTACKS ON NEURAL MODULATION RECOGNITION

Deep neural networks (DNNs) are widely used for neural modulation recognition (NMR) in the electronic field and have been shown to be vulnerable to adversarial examples for NMR. In the physical signal communication scenario, the adversarial signal transmitted by the adversary is affected by the channel, resulting in a random time delay with the original signal and causing decay on the attack performance. To address this issue, we propose the Position-Invariant adversarial attack Method (PIM) that generates the position-invariant adversarial signal by averaging the adversarial signals generated by shifted input signals to mitigate the channel effect on time delay. Our PIM can be easily integrated with other methods to achieve better results. Extensive experiments demonstrate that the proposed method could outperform all baselines for adversarial attacks on NMR under the time delay setting.