Copy, Right? A Testing Framework for Copyright Protection of Deep Learning Models

Explore a 19-minute IEEE conference talk on a testing framework for copyright protection of deep learning models. Delve into the DEEPJUDGE framework, multi-level testing metrics, and test case generation in both black-box and white-box settings. Examine experiments against model fine-tuning, pruning, extraction, and adaptive attackers. Gain insights into DNN watermarking and the challenges of protecting intellectual property in the field of artificial intelligence.

Syllabus

Intro
Deep Learning Models
DL Model Copyright Protection
Model Thief
DNN Watermarking
DEEPJUDGE Framework
Multi-level Testing Metrics
Metrics: Property-level
Metrics: Neuron-level
Metrics: Layer-level
Test Case Generation
Generation: Black-box Setting
Generation: White-box Setting
Final Judgement
Experiments
Against Model Fine-tuning & Pruning
Combined Visualization
Comparison
Against Model Extraction
Adaptive Attackers
Summary & Discussion