CorrGAN: Input Transformation Technique Against Natural Corruptions

TL;DR

CorrGAN is a novel GAN-based method that denoises natural corruptions in inputs, significantly improving DNN classification accuracy on corrupted data in real-time systems.

Contribution

Introduces CorrGAN, a GAN framework with a new loss function for denoising natural corruptions, enhancing DNN robustness against real-world input degradations.

Findings

Up to 75.2% of corrupted inputs correctly classified after denoising.

Effective against various natural corruptions like fog, blur, and contrast.

Demonstrates real-time applicability for safety-critical systems.

Abstract

Because of the increasing accuracy of Deep Neural Networks (DNNs) on different tasks, a lot of real times systems are utilizing DNNs. These DNNs are vulnerable to adversarial perturbations and corruptions. Specifically, natural corruptions like fog, blur, contrast etc can affect the prediction of DNN in an autonomous vehicle. In real time, these corruptions are needed to be detected and also the corrupted inputs are needed to be de-noised to be predicted correctly. In this work, we propose CorrGAN approach, which can generate benign input when a corrupted input is provided. In this framework, we train Generative Adversarial Network (GAN) with novel intermediate output-based loss function. The GAN can denoise the corrupted input and generate benign input. Through experimentation, we show that up to 75.2% of the corrupted misclassified inputs can be classified correctly by DNN using CorrGAN.