Novel Method to Detect Backdoor Attacks on Robust ML Models
Keeping your Backdoor Secure in your Robust Machine Learning Model
Tags: Singapore University of Technology and Design, Singapore, Computing Technology
SUTD researchers developed AEGIS, a pioneering technique to detect backdoor attacks in robust machine learning models, enhancing AI's security. AEGIS identifies backdoor-infected models by analyzing feature representations, a crucial step given the vulnerability of robust models to such attacks. Its application in image classifiers can significantly improve the trustworthiness of AI systems used in various sectors. AEGIS operates efficiently, detecting over 91% of backdoor attacks in minutes, setting a new standard in AI security. This innovation opens paths for safer AI applications, from self-driving cars to secure software systems.
IP Type or Form Factor: Software & Algorithm; Process & Method
TRL: Not specified
Industry or Tech Area: Computing Security; Computing Infrastructure & Networks
