Proj CDeepFuzz Paper Reading: Graph-based Fuzz Testing for Deep Learning Inference Engines

Abstract

本文：graph-based fuzz testing
Task: Fuzzing Machine Learning Libraries using operator graph
Github: https://github.com/gbftdlie/G (Q?https://github.com/gbftdlie/Graph-based-fuzz-testing)
Method:

1. Mutations: graph-based model-level mutations (graph edges addition, graph edges removal, block nodes addition, block nodes removal ) and source-level mutations (tensor shape mutation, parameter mutation)
2. operator-level coverage criterion
3. MCTS

实验：
对象: TensorFlow, 50 operators and 3 subgraphs
For RQ1, 1000 models are generated, For RQ2, RQ3 and RQ4, 400 test inputs are generated for each strategy
效果：

1. +40 exceptions, in three types: model conversion failure, inference failure, output comparison failure
2. 对mutation strategy: +8.2% operator level coverage, +8.6 exceptions