# What it Thinks is Important is Important: Robustness Transfers through Input Gradients

Source: [What it Thinks is Important is Important: Robustness Transfers through Input Gradients](<https://www.sota2.com/research/paper/68fa689e1e209131cea1d204>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-25T22:40:59.400Z

Authors: Alvin Chan, Yi Tay, Yew-Soon Ong

Published: 2019-12-11T23:51:37.000Z

## Abstract

Adversarial perturbations are imperceptible changes to input pixels that can change the prediction of deep learning models. Learned weights of models robust to such perturbations are previously found to be transferable across different tasks but this applies only if the model architecture for the source and target tasks is the same. Input gradients characterize how small changes at each input pixel affect the model output. Using only natural images, we show here that training a student model's input gradients to match those of a robust teacher model can gain robustness close to a strong baseline that is robustly trained from scratch. Through experiments in MNIST, CIFAR-10, CIFAR-100 and Tiny-ImageNet, we show that our proposed method, input gradient adversarial matching, can transfer robustness across different tasks and even across different model architectures. This demonstrates that directly targeting the semantics of input gradients is a feasible way towards adversarial robustness.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Adversarial Robustness Accuracy on CIFAR-100 (test) (FGSM and PGD Steps)](<https://www.sota2.com/research/sota/adversarial-robustness-accuracy-on-cifar-100-test-fgsm-and-pgd-steps.md>): Image Classification; CIFAR-100 (test); Clean Accuracy: 78.7.

- [Image Classification on CIFAR-10 (test) (Adversarial Robustness - PGD 100 & Auto Attack)](<https://www.sota2.com/research/sota/image-classification-on-cifar-10-test-adversarial-robustness-pgd-100-and-auto-attack.md>): Image Classification; CIFAR-10 (test); Clean Accuracy: 88.7.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/1912.05699>)

[PDF](<https://arxiv.org/pdf/1912.05699.pdf>)