Towards Defending Adversarial Attacks with Temperature Regularization in Automatic Modulation Recognition

Deep learning has been shown to perform extremely well at various machine learning tasks. However, these same architectures are highly vulnerable to adversarial examples: malicious inputs carefully crafted by adversaries which can force a neural network to produce erroneous predictions with high confidence. This undermines the security of deep learning algorithms when apply to those security-sensitive applications. Existing works have shown that the Signal Modulation Recognition (SMR) solutions based on deep learning are also susceptible to adversarial attacks. In this paper, we propose a new approach called temperature regularization to defense a deep learning scheme against white-box attacks in signal modulation recognition. Specifically, we introduce different temperatures to the softmax layer during the training of the neural network. Experimental results show that training a neural network with an appropriate high temperature can significantly enhance its robustness to three white-box attacks.