We propose using an adversarial autoencoder (AAE) to replace generativeadversarial network (GAN) in the private aggregation of teacher ensembles . The AAE architecture allows us to obtain good synthetic speech leveraging upona discriminative training of latent vectors . Such synthetic speech is used tobuild a privacy-preserving classifier when non-sensitive data is notsufficiently available in the public domain . The proposed PATE-AAE improves the average classification accuracy by +$2.11\%$ and +$6.60\%$, respectively, when compared with alternatives PATE and DP-GAN, while maintaining a strong level of privacy target at $0.01 with a fixed

Author(s) : Chao-Han Huck Yang, Sabato Marco Siniscalchi, Chin-Hui Lee

Links : PDF - Abstract

Code :
Coursera

Keywords : aae - pate - privacy - aggregation - speech -

Leave a Reply

Your email address will not be published. Required fields are marked *