Adversarial Training with Orthogonal Regularization

2020-10-05

SIU 2020Signal Processing and Communications Applications Conference

Adversarial Training with Orthogonal Regularization

Oğuz Kaan Yüksel·İnci Meliha Baytaş
October 2020
Paper
We propose orthogonal regularization during adversarial training to improve robustness. By encouraging weight matrices to stay orthogonal, the model learns more stable and transferable representations under attack.

Research context

Open in graph

Addresses

Adversarial robustness

Uses

Adversarial training Orthogonality penalty

Studies

Generalization Learned representation

Produces

AlgorithmAdversarial training

Orthogonally regularized adversarial training

An adversarial-training algorithm augmented with an orthogonality penalty to promote stable and transferable representations.

Related publications