A Tool to Detect Privacy Leaks in Deep Learning Models

About the Technology

As machine learning models are widely deployed across APIs, they become vulnerable to adversarial threats. Specifically, Model Extraction or Model Stealing attacks allow adversaries to query the API systematically and train a duplicate copy of the model, violating intellectual property and security.

This research aims to develop an assessment tool that allows end-user agencies to audit ML models before deploying them in production. By querying the API model internally, the tool simulates extraction attacks to quantify vulnerabilities and ensure informative deployment decisions.

Novel Features

Unlike complex, ad-hoc vulnerability testing methods, this tool abstracts the process into a single, unified configuration file. The user specifies evaluation parameters in a simple text file, and the tool automatically executes security benchmarks, generating a structured report detailing potential privacy leaks and intellectual property vulnerabilities.

National & Global Impact

Ensuring cybersecurity in national artificial intelligence systems is critical. This tool provides public agencies and private enterprises with a standardized, automated cybersecurity auditing system for deep learning models. The library is distributed openly via PyPI, allowing developers globally to test and fortify their API-based models.