DNN-GP: Diagnosing and Mitigating Model's Faults Using Latent Concepts

Watch a research presentation from USENIX Security '24 that introduces DNN-GP, an innovative fault diagnosis tool for Deep Neural Networks. Explore how this integrated interpreter diagnoses various model faults through latent concept interpretation, addressing critical issues in DNN robustness and concept drift. Learn about the tool's unique approach of using probing samples from adversarial attacks, semantic attacks, and drift samples to interpret erroneous model decisions. Discover how DNN-GP develops countermeasures in concept space to enhance model resilience, featuring transferable training capabilities for unsupervised diagnosis across different models. See the tool's impressive performance demonstrated across three real-world datasets, achieving nearly 100% detection accuracy while maintaining low false positive rates.