A Detection Algorithm for Audio Adversarial Examples in EI-Enhanced Automatic Speech Recognition

Benefiting from the development of big data, edge computing, and deep learning, splendid breakthroughs have been made in automatic speech recognition (ASR) recent years. Since then, more smart products chosen as interface for human-computer interaction, which causes popularity intelligence (EI) enhanced recognition. While people are enjoying social changes brought by technology, a factor instability quietly emerged called audio adversarial example is type deliberately generated attackers via adding subtle perturbations to original signal. The added sound like certain noise that cannot be precepted human but will cause ASR system make wrong transcription. Three detection algorithms examples proposed this thesis, namely, robust algorithm based on WER (word error rate), feature ADR (adversarial ratio), collaborative neural network. experiment results show three thesis great discrimination achieve high AUC scores. Among them, cooperative best worst. In addition, we found tends higher accuracy score lower recall score, while converse performance. Moreover, since method combines advantages methods, it presents better performance with respect accuracy, recall, F1 score.