Inverting Neural Networks: New Methods to Generate Neural Network Inputs from Prescribed Outputs

TL;DR

This paper introduces two novel methods for inverting neural networks to generate input images from specific outputs, revealing vulnerabilities and providing deeper understanding of network decision boundaries.

Contribution

The paper presents two new general methods for solving the inverse problem in neural networks, enabling the generation of input images from prescribed outputs.

Findings

Methods produce random-like images with high classification accuracy

Reveals vulnerabilities in neural network architectures

Applicable to transformer and linear layer networks

Abstract

Neural network systems describe complex mappings that can be very difficult to understand. In this paper, we study the inverse problem of determining the input images that get mapped to specific neural network classes. Ultimately, we expect that these images contain recognizable features that are associated with their corresponding class classifications. We introduce two general methods for solving the inverse problem. In our forward pass method, we develop an inverse method based on a root-finding algorithm and the Jacobian with respect to the input image. In our backward pass method, we iteratively invert each layer, at the top. During the inversion process, we add random vectors sampled from the null-space of each linear layer. We demonstrate our new methods on both transformer architectures and sequential networks based on linear layers. Unlike previous methods, we show that our new methods are able to produce random-like input images that yield near perfect classification scores in all cases, revealing vulnerabilities in the underlying networks. Hence, we conclude that the proposed methods provide a more comprehensive coverage of the input image spaces that solve the inverse mapping problem.