Pith. sign in

REVIEW

Verification of Recurrent Neural Networks Through Rule Extraction

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 1811.06029 v1 pith:TAW6KPYS submitted 2018-11-14 cs.LG stat.ML

Verification of Recurrent Neural Networks Through Rule Extraction

classification cs.LG stat.ML
keywords networksrecurrentadversarialperturbationgrammarsneuralsamplesverification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The verification problem for neural networks is verifying whether a neural network will suffer from adversarial samples, or approximating the maximal allowed scale of adversarial perturbation that can be endured. While most prior work contributes to verifying feed-forward networks, little has been explored for verifying recurrent networks. This is due to the existence of a more rigorous constraint on the perturbation space for sequential data, and the lack of a proper metric for measuring the perturbation. In this work, we address these challenges by proposing a metric which measures the distance between strings, and use deterministic finite automata (DFA) to represent a rigorous oracle which examines if the generated adversarial samples violate certain constraints on a perturbation. More specifically, we empirically show that certain recurrent networks allow relatively stable DFA extraction. As such, DFAs extracted from these recurrent networks can serve as a surrogate oracle for when the ground truth DFA is unknown. We apply our verification mechanism to several widely used recurrent networks on a set of the Tomita grammars. The results demonstrate that only a few models remain robust against adversarial samples. In addition, we show that for grammars with different levels of complexity, there is also a difference in the difficulty of robust learning of these grammars.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.