Responsible AI for Offline Plugins
Tamper-Resistant Neural Audio Watermarking
In the era of rapid generative AI advancements, how do audio plugin developers harness the power of deep learning, while safeguarding against misuses and ensuring regulatory compliance? Existing audio watermarking algorithms are designed to run on the cloud, and when deployed offline, are trivial to bypass when faced against experienced hackers. In this talk, I introduce an innovative algorithm developed for Vocoflex, our AI powered voice transformation plugin. Instead of first generating the voice and then adding the watermark in a post-processing step, our approach involves an end-to-end trained neural network that generates voices with inaudible watermarks already embedded. Furthermore, the neural network resists tampering by 'self-destructing' to contaminate outputs upon malicious perturbation of the network weights. We will share our methodology and provide guidelines for designing similar, but not identical watermarking schemes that tightly couple program logic with watermarks, resilient to both audio manipulations and reverse engineering attempts.
Kanru Hua
CEO
Dreamtonics
