Authenticating: Deep Generative Models

Mayank Kumar Singh, Naoya Takahashi, Wei-Hsiang Liao, Yuki Mitsufuji
{mayank.a.singh, naoya.takahashi, weihsiang.liao, yuki.mitsufuji}@sony.com

Abstract

This paper introduces a novel approach to deter unauthorized use and facilitate user tracking in generative models by integrating key-based authentication with watermarking techniques. Our method involves providing users with model parameters accompanied by a unique, user-specific key. During inference, the model requires both the key and the standard input. If the key is valid, the model produces the expected output; otherwise, it generates a degraded output, thereby enforcing key-based model authentication. To enable user tracking, the model embeds the user's unique key as a watermark within the generated content, allowing for the identification of the user's unique ID within the metadata of the generated output. We demonstrate the effectiveness of our approach on two types of models—audio codecs and vocoders—utilizing the SilentCipher watermarking method. Additionally, we assess the robustness of the embedded watermarks against various distortions, validating their reliability in different scenarios.

Audio Samples

Please use wired studio quality headphones to evaluate the samples.
Listen to the samples at a higher loudness than your normal settings.

Encodec-32khz Model
Original
SilentCipher Watermarked
Valid Key
Invalid Key





HiFiGAN Model
Original
SilentCipher Watermarked
Valid Key
Invalid Key