Multidiscriminator Sobolev Defense-GAN Against Adversarial Attacks for End-to-End Speech Systems

This paper introduces a defense approach against end-to-end adversarial attacks developed for cutting-edge speech-to-text systems. The proposed algorithm has four steps. First, we use the short-time Fourier transform to represent speech signals with 2D spectrograms. Second, iteratively find safe vector using spectrogram subspace projection operation. operation minimizes chordal distance adjustment between spectrograms an additional regularization term. Third, synthesize such novel GAN architecture trained Sobolev integral probability metric. We impose constraint on generator network improve model’s performance in terms of stability and total number learned modes. Finally, reconstruct signal from synthesized Griffin-Lim phase approximation technique. evaluate six strong white black-box DeepSpeech, Kaldi, Lingvo models. experimental results show that our outperforms other state-of-the-art algorithms accuracy quality.