Research on Adversarial Sample Attack Defense Methods for Speech Recognition

Recently, deep learning has achieved significant success in various fields such as image classification, speech recognition, and natural language understanding. However, the security issues associated with deep learning are becoming increasingly apparent, particularly the challenges posed byad-versarial sample attacks. These attacks, which are both covert and potentially harmful, pose substantial security threats to deep learning models. In response to this, we propose a novel defense method against adversarial sample attacks tailored for speech recognition. This method involves constructing an audio reconstruction model that restores adversarial speech samples back to their normal state. The model incorporates an im-proved Wavenet residual network structure based on dilated convolutions, and utilizes Mel Frequency Cepstral Coefficients to enhance audio reconstruction. Experimental results indicate that the proposed audio reconstruction model exhibits superior reconstruction capabilities, effectively restoring and defending against various adversarial sample attacks.