A Mathematical Proof of Parallel Composition for Approximate Differential Privacy
A Mathematical Proof of Parallel Composition for Approximate Differential Privacy
[Read More]
-
-
Intellectual Property Protection of Speaker Recognition ML Models
Problem Statement In recent years, research has increased on model protection techniques, especially in the image domain. On the other hand, the audio domain, specifically speaker recognition (SR) models and ways to protect them from being stolen, is still rather unexplored. Speaker recognition models are designed to identify and verify... [Read More] -
Record Linkage Using Metagenomic Microbiome Profiles
Problem Statement In healthcare, record linkage enables researchers to measure cross-sector care, assess care integration, and consider long-term outcomes from different sources. Deterministic linkage is applied when one or several identifiers can be used alone or in combination. However, these identifiers are not always reliable, as linkage errors can arise... [Read More] -
Securing DNN models deployed on edge devices with obfuscation
Introduction The Obfuscation is a form of security via obscurity. In this context, obfuscation of DNN models deployed on edge devices such as mobile phones, computers, cameras, automotive gadgets, and any user-owned devices. If the user who owns the device on which the DNN model and code are copied locally,... [Read More] -
Machine Learning Fingerprinting
Enhancing Deep Neural Network Security
Fingerprinting is the process of embedding unique markers within a model to verify its authenticity and integrity. It not only facilitates the distinction between different users of the same model, enabling traceability and accountability, it also plays a crucial role in ensuring the secure deployment of machine learning models through... [Read More]