Towards verification of Natural Language Processing models
Abstract:
In recent years, deep neural networks (DNN) have become ubiquitous in our lives. Many efforts have been made to assess the behavior of these networks, especially their robustness against adversarial attacks. Going one step further, researchers have developed verification techniques to prove that certain properties hold, meaning that the system fulfills specific requirements. Nonetheless, DNN verification has been mainly applied in the computer vision domain. When applied to Natural Language Processing (NLP), current techniques cannot solve many verification problems. In this work, we introduce VeriPipe, an extended verification pipeline with an additional module PCV, that uses a novel property specification for NLP networks and can solve more problems than state-of-the-art NLP verification techniques, by tightening the input constraints. We evaluated VeriPipe using different verification tools to showcase its applicability, and our results show that it improves the results produced by state-of-the-art NLP verification.
Committee:
- Matthew Dwyer, Committee Chair, CS/SEAS/UVA
- Sebastian Elbaum, Advisor, CS/SEAS/UVA
- Yangfeng Ji, CS/SEAS/UVA
- Thomas Fletcher, CS/ECE/SEAS/UVA