LEMNA - Explaining Deep Learning Based Security Applications

Explore a 21-minute conference talk that delves into LEMNA, a novel approach for explaining deep learning-based security applications. Learn about the challenges of opaque deep learning models in security-critical domains and the limitations of existing explanation techniques. Discover how LEMNA addresses these issues by supporting locally non-linear decision boundaries and modeling feature dependency. Gain insights into deriving explanations from deep neural networks, evaluating explanation accuracy, and practical applications such as identifying binary function starts. Understand how LEMNA contributes to building trust in target models and aids in troubleshooting and patching model errors, ultimately enhancing the transparency and reliability of deep learning in security contexts.

Intro
The Concerns of Opaque Deep Learning Model
Existing Explanation Techniques & Limitations
One Example of Model Explanation (LIME. KDD'16)
Limitation of Existing Explanation Techniques
LEMNA: Local Explanation Method using Nonlinear Approximation
Supporting Locally Non-linear Decision Boundaries
Modeling the Feature Dependency . Mature regression model with fused lasso
Deriving an Explanation from DNN with LEMNA
Explanation Accuracy Evaluation
Demonstration of LEMNA in Identifying Binary Function Start
Building Trust in the Target Models
Troubleshooting and Patching Model Errors