ALFASVMLib is an open-source Matlab library that implements a set of heuristic attacks against Support Vector Machines (SVMs). The goal of such attacks is to maximally compromise the SVM's classification accuracy by mislabeling a given fraction of training samples. They are indeed referred to as adversarial label flip attacks, and can be also considered a form of worst-case label noise.

For updated code and documentation, please refer to our repository on GitHub.

Details of each attack strategy can be found in our paper:

H. XiaoB. BiggioB. NelsonH. XiaoC. Eckert, and F. RoliSupport Vector Machines under Adversarial Label ContaminationNeurocomputing, vol. Special Issue on Advances in Learning with Label Noise, In Press. BibTex  (2.8 MB)

ALFASVMLib is free software, released under the GNU GPL v3 license.