Responsible AI for Offline Plugins - Tamper-Resistant Neural Audio Watermarking

This conference talk from the Audio Developer Conference (ADC) 2024 explores innovative approaches to neural audio watermarking for offline AI plugins. Learn how Kanru Hua, founder of Tokyo-based audio software manufacturer Dreamtonics, addresses the challenges of safeguarding AI-powered audio tools against misuse while ensuring regulatory compliance. Discover a groundbreaking algorithm developed for Vocoflex that integrates inaudible watermarks directly into the voice generation process through end-to-end trained neural networks, rather than applying watermarks as a post-processing step. The presentation details how these networks can "self-destruct" to contaminate outputs when malicious tampering is detected, creating a system resistant to both audio manipulations and reverse engineering attempts. Gain valuable insights into designing watermarking schemes that tightly couple program logic with watermarks for enhanced security in offline AI audio applications. The 22-minute talk includes access to presentation slides and represents part of ADC's mission to advance audio development technologies across music applications, game audio, audio processing, and embedded systems.