Deep Forgery Defense System for Live ‎Agricultural E-Commerce—A ‎Meta-Learning-Based Dynamic Verification ‎Model for Anchor Voiceprints

Abstract

With the booming development of agricultural products live e-commerce, the threat of deep forgery technology is ‎becoming more and more prominent. In this paper, a deep forgery defense system for live agricultural products ‎e-commerce is constructed to guarantee the authenticity and security of e-commerce. Based on MFCC features for voice ‎feature extraction, and for the defects of ObamaNet synthesized anchor baseline system with large time overhead, a ‎speech synthesis model for live agricultural products e-commerce based on mouth shape classification is proposed to ‎match voice features with mouth shape. In addition, a semi-supervised voice verification model is proposed based on ‎meta-learning, in which the meta-gradient computed by the model on the training set is used as a guide for gradient ‎optimization on the test set, so that the model learns richer knowledge. The experiments show that there is a big difference ‎between the synthesized speech and the prototype speech in terms of voiceprint features, and some syllables also have ‎significant differences in the interval of 3kHz~4kHz, which indicates that the system in this paper can effectively ‎distinguish the real speech from the forged speech, and enhances the reliability of the system, which plays an important ‎role in the field of live e-commerce of agricultural products.‎