Publications#
The underlying techniques used in DNNV are described in several publications. Additionally, DNNV has been used by several works to facilitate the implementation of new techniques, as well as the usage and comparison of many verification tools on novel problems.
Property reduction for DNN properties was first introduced in Reducing DNN Properties to Enable Falsification with Adversarial Attacks.
The DNNV tool is introduced in DNNV: A Framework for Deep Neural Network Verification, along with overviews of the DNN simplifications and property reductions used.
The paper Systematic Generation of Diverse Benchmarks for DNN Verification used DNNV to quickly run many verification tools on the benchmarks generated by their technique.
The paper Distribution Models for Falsification and Verification of DNNs used DNNV to evaluate their approach for specifying properties over complex input domains, such as images.