ENHANCING MODEL SECURITY - LEVERAGING USER-GENERATED IDS AS EMBEDDED WATERMARKS IN MACHINE LEARNING MODELS

Aug 3, 2026

Machine learning model theft or unauthorized reuse can cause substantial financial losses and serious damage to a company’s intellectual property. Existing protections such as encryption or access controls can be bypassed by skilled adversaries. The document proposes embedding embedded watermarks directly into machine learning models, enabling unique model identification and user-level identification. If a model is leaked or misused, watermark extraction supports source traceability and accountability, and helps enable detection and prosecution of unauthorized use.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

AccessFree to download

Abstract

Detailed Description

Threat and Limitations of Existing Protections

Embedded Watermarking Approach

Watermark Encoding, Embedding, and Recovery

Example UserID Watermarking Operations

Technical Disclosure Commons Defensive Publications Series October 2023 ENHANCING MODEL SECURITY: LEVERAGING USER-GENERATED IDS AS EMBEDDED WATERMARKS IN MACHINE LEARNING MODELS Alan Gatzke Mitchell C Mosure Andi Wilson Ananth Racherla Todd C Kuehnl See next page for additional authors Follow this and additional works at: [https://www.tdcommons.org/dpubs_series](https://www.tdcommons.org/dpubs_series) Recommended Citation Gatzke, Alan; C Mosure, Mitchell; Wilson, Andi; Racherla, Ananth; C Kuehnl, Todd; and McCarthy, Tyler Shane, "ENHANCING MODEL SECURITY: LEVERAGING USER-GENERATED IDS AS EMBEDDED WATERMARKS IN MACHINE LEARNING MODELS", Technical Disclosure Commons,(October 23, 2023) [https://www.tdcommons.org/dpubs_series/6337](https://www.tdcommons.org/dpubs_series/6337) This work is licensed under a Creative Commons Attribution 4.0 License. This Article is brought to you for free and open access by Technical Disclosure Commons. It has been accepted for inclusion in Defensive Publications Series by an authorized administrator of Technical Disclosure Commons. Inventor(s) Alan Gatzke, Mitchell C Mosure, Andi Wilson, Ananth Racherla, Todd C Kuehnl, and Tyler Shane McCarthy This article is available at Technical Disclosure Commons: [https://www.tdcommons.org/dpubs_series/6337](https://www.tdcommons.org/dpubs_series/6337) ENHANCING MODEL SECURITY: LEVERAGING USER-GENERATED IDS AS EMBEDDED WATERMARKS IN MACHINE LEARNING MODELS AUTHORS: Alan Gatzke Mitchell C Mosure Andi Wilson Ananth Racherla Todd C Kuehnl Tyler Shane McCarthy ABSTRACT The potential theft or unauthorized use of machine learning models developed by a company can lead to significant financial losses and damage to the company's intellectual property. While existing methods of protecting machine learning models such as encryption or access controls can be circumvented by skilled attacker, techniques presented herein involve the integration of embedded watermarks into machine learning models. Such techniques involving the integration of embedded watermarks may not only uniquely identify a model but may also include a unique user identification/identity that can make it possible to track usage of the model and detect any unauthorized use of the model. Thus, if a model is leaked, redistributed, or misused, the watermark for the model makes it possible to identify a source of the leak/misuse, allowing for better traceability and accountability. DETAILED DESCRIPTION Proposed herein are techniques to address the potential theft or unauthorized use of machine learning models, which can lead to significant financial losses and damage to acompany's intellectual property. Many existing methods of protecting machine learning models, such as encryption or access controls, can be circumvented by skilled attackers. However, techniques presented herein involve the integration of embedded watermarks into machine learning models. These watermarks not only uniquely identify a model but can also include a unique user identification, enabling the tracking of model usage and detection of any unauthorized use. By implementing embedded watermarks with unique user identifications, the techniques proposed herein can significantly reduce the risk of 1 6966 Published by Technical Disclosure Commons, 2023 2 intellectual property theft and also facilitate the detection and prosecution of any unauthorized use. In accordance with techniques of this proposal, a watermarking system can inject user personalized watermarks in a model at runtime of the model. Broadly, operations of the watermarking system may include three high-level steps involving: 1) encoding a watermark, which is a unique identifier/identity (ID) tied to a specific user, into a form that can be hidden in the cover data of a model (e.g., a sparse pattern or algorithm could be used to change the watermark information into chosen parts of a model); 2) embedding the

FAQs

Why are traditional protections like encryption or access controls insufficient?

They can be circumvented by skilled attackers, enabling theft or unauthorized use of machine learning models.

How do embedded watermarks help protect machine learning models?

They uniquely identify a model and can include a user identity to track usage and detect unauthorized use.

What are the main steps of the proposed watermarking system?

Encode a user-tied ID into a hidden form, embed it into selected model cover data (weights or biases) with minimal performance impact, then recover the ID by extracting the hidden watermark via decoding or a trained helper model.

Aug 3, 2026

Upload by Ava Thompson