Home

Overview

TensorFI is a configurable fault injector for applications written using the TensorFlow framework. Unlike existing fault injectors such as LLFI, FSEFI, PINFI, etc., TensorFI injects high-level faults directly into TensorFlow programs. As TensorFlow programs are represented as dataflow graphs, users inject faults at the level of graph nodes. This ensures that programmers can establish a one-to-one correspondence between the injected fault and the effect of the fault on the machine learning algorithm.

The goal of the following pages is to provide information on the implementation, usage and programming of the TensorFI tool.

Design Principles

How TensorFI works