On the unreasonable effectiveness of CNNs

TL;DR

This paper demonstrates that standard CNN architectures like U-Net can effectively solve complex inverse problems such as XOR decryption from noisy data, highlighting their broad applicability in image reconstruction tasks.

Contribution

It provides empirical evidence of CNNs' effectiveness on a novel inverse problem, extending their known capabilities beyond traditional imaging applications.

Findings

U-Net achieves acceptable results on XOR decryption from noisy data

CNNs can be applied to complex inverse problems in cryptography

Standard CNN architectures have broad potential in image reconstruction

Abstract

Deep learning methods using convolutional neural networks (CNN) have been successfully applied to virtually all imaging problems, and particularly in image reconstruction tasks with ill-posed and complicated imaging models. In an attempt to put upper bounds on the capability of baseline CNNs for solving image-to-image problems we applied a widely used standard off-the-shelf network architecture (U-Net) to the "inverse problem" of XOR decryption from noisy data and show acceptable results.