PORTALE DELLA DIDATTICA

Ricerca CERCA
  KEYWORD

Image Processing Lab (IPL)

Adversarial learning for robust multiclass classification by learning mappings onto target Gaussian distributions

Parole chiave DEEP LEARNING, DEEP NEURAL NETWORKS, MACHINE LEARNING, MACHINE LEARNING, ARTIFICIAL NEURAL NETWORKS

Riferimenti ENRICO MAGLI

Gruppi di ricerca CCNE - COMMUNICATIONS AND COMPUTER NETWORKS ENGINEERING, ICT4SS - ICT FOR SMART SOCIETIES, Image Processing Lab (IPL)

Tipo tesi RESEARCH

Descrizione Adversarial models are now the de-facto standard to approach generative models. The first generative model trained by means of an adversarial loss, the Generative Adversarial Network (GAN) [1], gained immediate popularity and opened the path to the field of adversarial training. GANs represent a shift in architecture design for deep neural networks. This new architecture puts two or more neural networks against each other in adversarial training to produce generative models.

It has been established in AuthNet [2] that the adversarial models can be effectively used for regularizing the latent space on target Gaussian distributions that lead to robust authentication. Since this system is developed for two classes, authorized and non-authorized, the purpose of this thesis is to expand it to a multi-class system and analyze the results.

Tasks:

* Initial study for three classes: Study of the latent space and gaussianity of the distributions.

* Study of optimal target mean and variance

* Introduction of higher order moments in the loss function for improving the accuracy and regularization of the distributions

* Comparison of results with standard Cross Entropy training

* Study of robustness against Adversarial attacks models of FGSM, TGSM, JSMA, and PGD

* Employing Adversarial training and defensive distillation for improving robustness

References:

[1] Goodfellow, Ian, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. "Generative adversarial nets." In Advances in neural information processing systems, pp. 2672-2680. 2014.


[2] Ali, Arslan, Matteo Testa, Tiziano Bianchi, and Enrico Magli. "Authnet: Biometric Authentication Through Adversarial Learning." In 2019 IEEE 29th International Workshop on Machine Learning for Signal Processing (MLSP), pp. 1-6. IEEE, 2019.

Conoscenze richieste Candidate students should have some background on neural networks. Some experience of TensorFlow environment and Python programming are desirable, along with good programming skills.


Scadenza validita proposta 17/01/2023      PROPONI LA TUA CANDIDATURA




© Politecnico di Torino
Corso Duca degli Abruzzi, 24 - 10129 Torino, ITALY
Contatti