Towards Efficient and Effective Adversarial Training
6. prosince 2021
Řečníci
Gaurang Sriramanan
Řečník · 1 sledující
Sravanti Addepalli
Řečník · 0 sledujících
Arya Baburaj
Řečník · 0 sledujících
R. Venkatesh Babu
Řečník · 0 sledujících
O prezentaci
The vulnerability of Deep Neural Networks to adversarial perturbations has spurred immense interest towards the development of models that are robust to such attacks. However, present state-of-the-art adversarial defenses involve the use of 10-step adversaries, which renders them computationally infeasible for application to large-scale datasets. While recent single-step adversarial training methods show promising direction, their robustness is not on par with multi-step training methods. In this work, we bridge this performance gap by introducing a novel Nuclear-Norm regularizer on network predictions to enforce function smoothing in the vicinity of data samples, and boost the accuracy further by using Weight Averaging for both attack generation and training. While prior works consider each data sample independently, the proposed Nuclear-Norm Adversarial Training (NuAT) imposes a rank minimization constraint on the oscillation of function values across a training minibatch, thereby enhancing optimization using the joint batch-statistics of adversarial samples. We achieve state-of-the-art results amongst existing efficient training methods and demonstrate results which are comparable to that of multi-step adversarial training methods such as TRADES and PGD-AT, at a significantly lower computational cost.The vulnerability of Deep Neural Networks to adversarial perturbations has spurred immense interest towards the development of models that are robust to such attacks. However, present state-of-the-art adversarial defenses involve the use of 10-step adversaries, which renders them computationally infeasible for application to large-scale datasets. While recent single-step adversarial training methods show promising direction, their robustness is not on par with multi-step training methods. In thi…
Organizátor
NeurIPS 2021
Účet · 1,9k sledujících
O organizátorovi (NeurIPS 2021)
Neural Information Processing Systems (NeurIPS) is a multi-track machine learning and computational neuroscience conference that includes invited talks, demonstrations, symposia and oral and poster presentations of refereed papers. Following the conference, there are workshops which provide a less formal setting.
Baví vás formát? Nechte SlidesLive zachytit svou akci!
Profesionální natáčení a streamování po celém světě.
Sdílení
Doporučená videa
Prezentace na podobné téma, kategorii nebo přednášejícího
Machine learning in practice: Who is benefiting? Who is being harmed?
Zhlédnout později
Timnit Gebru, …
Pro uložení prezentace do věčného trezoru hlasovalo 0 diváků, což je 0.0 %
PreferenceNet: Encoding Human Preferences in Auction Design with Deep Learning
Zhlédnout později
Neehar Peri, …
Pro uložení prezentace do věčného trezoru hlasovalo 0 diváků, což je 0.0 %
A Toolbox for Construction and Analysis of Speech Datasets
Zhlédnout později
Pro uložení prezentace do věčného trezoru hlasovalo 0 diváků, což je 0.0 %
A Gradient Method for Multilevel Optimization
Zhlédnout později
Ryo Sato, …
Pro uložení prezentace do věčného trezoru hlasovalo 0 diváků, což je 0.0 %
(Almost) Free Incentivized Exploration from Decentralized Learning Agents
Zhlédnout později
Pro uložení prezentace do věčného trezoru hlasovalo 0 diváků, což je 0.0 %
Equilibrium and non-Equilibrium regimes in the learning of Restricted Boltzmann Machines
Zhlédnout později
Pro uložení prezentace do věčného trezoru hlasovalo 0 diváků, což je 0.0 %