Date of Award

6-2-2026

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2026

Department

Computer Science and Engineering

First Advisor

Yi Fang

Abstract

Hemlock is a tool designed to protect musicians from having their work used to train generative AI models without their consent. It works by adding carefully crafted inaudible noise to audio files that disrupts the ability of AI models to learn from them — a technique known as adversarial perturbation. Our system targets three different types of AI model simultaneously: a Music Information Retrieval (MIR) classifier, a sequential audio generation model (Mel-LSTM), and Meta’s AudioCraft, a transformer-based music generator. Testing in twenty songs showed an average 15% reduction in the MIR model’s classification confidence, a 59% increase in Mel- LSTM prediction error, and a 9% decrease in AudioCraft’s cosine similarity. The perturbations also transferred to Wav2Vec, a model against which the system was never trained, suggesting a broader generalizability. The algorithm is delivered as a VST3 plugin that integrates directly into digital audio workstations, making it accessible to working musicians and producers without requiring any technical expertise.

Download

Share

COinS