Robust machine learning typically refers to the robustness of machine learning algorithms. For a machine learning algorithm to be considered robust, either the testing error has to be consistent with the training error, or the performance is stable after adding some noise to the dataset.<ref>{{cite web |author=El Sayed Mahmoud |url=https://www.researchgate.net/post/What_is_the_definition_of_the_robustness_of_a_machine_learning_algorithm |title=What is the definition of the robustness of a machine learning algorithm? |url=https://www.researchgate.net/post/What_is_the_definition_of_the_robustness_of_a_machine_learning_algorithm |access-date=2016-11-13}}</ref> Recently, consistently with their rise in popularity, there has been an increasing interest in the robustness of neural networks. This is particularly due their vulnerability to adverserial attacks.<ref>{{cite journalarXiv |last1=Li |first1=Linyi |last2=Xie |first2=Tao |last3=Li |first3=Bo |title=SoK: Certified Robustness for Deep Neural Networks |journalarxiv=arXiv:2009.04131 [cs, stat] |date=9 September 2022}} {{DOI|doi=https://doi.org/10.48550/arXiv.2009.04131 |url=https://arxiv.org/pdf/2009.04131.pdf}}.</ref>
===Robust network design===
Line 38:
=== Robust algorithms ===
There exists algorithms that tolerate errors in the input<ref>{{cite book |last1=Carbin |first1=Michael |title=Proceedings of the 19th international symposium on Software testing and analysis - ISSTA '10 |last2=Rinard |first2=Martin C. |chapter=Automatically identifying critical input regions and code in applications |date=12 July 2010 |pages=37–48 |doi=10.1145/1831708.1831713 |publisher=ACM |isbn=9781605588230 |s2cid=1147058 |chapter-url=http://people.csail.mit.edu/rinard/paper/issta10.pdf |isbn=9781605588230 |s2cid=1147058 }}</ref> or during the computation.<ref name="Danglot">{{cite journal |last1=Danglot |first1=Benjamin |last2=Preux |first2=Philippe |last3=Baudry |first3=Benoit |last4=Monperrus |first4=Martin |title=Correctness attraction: a study of stability of software behavior under runtime perturbation |journal=Empirical Software Engineering |date=21 December 2017 |volume=23 |issue=4 |pages=2086–2119 |doi=10.1007/s10664-017-9571-8 |url=https://hal.archives-ouvertes.fr/hal-01378523/document |arxiv=1611.09187 |s2cid=12549038 }}</ref> In that case, the computation eventually converges to the correct output. This phenomenon has been called "correctness attraction".<ref name="Danglot"/>
