Publications by authors named "Loic Shi-Garrier"
Entropy (Basel)
January 2024
Article Synopsis
- - Despite their success, deep learning models struggle with robustness, especially against adversarial attacks, which raises concerns about their reliability in critical applications.
- - The paper proposes an information geometric framework to set clear robustness standards for l2 white-box attacks in multi-class classifications, using the Fisher information metric and specific criteria for the input-output Jacobian.
- - The proposed method shows significant enhancements in model robustness on MNIST and CIFAR-10 datasets against various adversarial attacks, outperforming existing techniques like defensive distillation and Jacobian regularization, while still achieving high accuracy.
