Timezone: »
Recent work has uncovered the interesting (and somewhat surprising) finding that training models to be invariant to adversarial perturbations requires substantially larger datasets than those required for standard classification. This result is a key hurdle in the deployment of robust machine learning models in many real world applications where labeled data is expensive. Our main insight is that unlabeled data can be a competitive alternative to labeled data for training adversarially robust models. Theoretically, we show that in a simple statistical setting, the sample complexity for learning an adversarially robust model from unlabeled data matches the fully supervised case up to constant factors. On standard datasets like CIFAR- 10, a simple Unsupervised Adversarial Training (UAT) approach using unlabeled data improves robust accuracy by 21.7% over using 4K supervised examples alone, and captures over 95% of the improvement from the same number of labeled examples. Finally, we report an improvement of 4% over the previous state-of-the- art on CIFAR-10 against the strongest known attack by using additional unlabeled data from the uncurated 80 Million Tiny Images dataset. This demonstrates that our finding extends as well to the more realistic case where unlabeled data is also uncurated, therefore opening a new avenue for improving adversarial training.
Author Information
Jean-Baptiste Alayrac (Deepmind)
Jonathan Uesato (DeepMind)
Po-Sen Huang (DeepMind)
Alhussein Fawzi (DeepMind)
Robert Stanforth (DeepMind)
Pushmeet Kohli (DeepMind)
More from the Same Authors
-
2021 Spotlight: Make Sure You're Unsure: A Framework for Verifying Probabilistic Specifications »
Leonard Berrada · Sumanth Dathathri · Krishnamurthy Dvijotham · Robert Stanforth · Rudy Bunel · Jonathan Uesato · Sven Gowal · M. Pawan Kumar -
2021 : Inferring a Continuous Distribution of Atom Coordinates from Cryo-EM Images using VAEs »
Dan Rosenbaum · Marta Garnelo · Michal Zielinski · Charles Beattie · Ellen Clancy · Andrea Huber · Pushmeet Kohli · Andrew Senior · John Jumper · Carl Doersch · S. M. Ali Eslami · Olaf Ronneberger · Jonas Adler -
2022 : Solving Math Word Problems with Process-based and Outcome-based Feedback »
Jonathan Uesato · Nate Kushman · Ramana Kumar · H. Francis Song · Noah Siegel · Lisa Wang · Antonia Creswell · Geoffrey Irving · Irina Higgins -
2022 : Panel »
Jeevana Priya Inala · Pushmeet Kohli · Ann Kennedy · Sriram Rajamani · Yisong Yue -
2022 : Solving Math Word Problems with Process-based and Outcome-based Feedback »
Jonathan Uesato · Nate Kushman · Ramana Kumar · H. Francis Song · Noah Siegel · Lisa Wang · Antonia Creswell · Geoffrey Irving · Irina Higgins -
2022 Poster: Flamingo: a Visual Language Model for Few-Shot Learning »
Jean-Baptiste Alayrac · Jeff Donahue · Pauline Luc · Antoine Miech · Iain Barr · Yana Hasson · Karel Lenc · Arthur Mensch · Katherine Millican · Malcolm Reynolds · Roman Ring · Eliza Rutherford · Serkan Cabi · Tengda Han · Zhitao Gong · Sina Samangooei · Marianne Monteiro · Jacob L Menick · Sebastian Borgeaud · Andy Brock · Aida Nematzadeh · Sahand Sharifzadeh · Mikołaj Bińkowski · Ricardo Barreira · Oriol Vinyals · Andrew Zisserman · Karén Simonyan -
2022 Poster: Characteristics of Harmful Text: Towards Rigorous Benchmarking of Language Models »
Maribeth Rauh · John Mellor · Jonathan Uesato · Po-Sen Huang · Johannes Welbl · Laura Weidinger · Sumanth Dathathri · Amelia Glaese · Geoffrey Irving · Iason Gabriel · William Isaac · Lisa Anne Hendricks -
2021 : Inferring a Continuous Distribution of Atom Coordinates from Cryo-EM Images using VAEs »
Dan Rosenbaum · Marta Garnelo · Michal Zielinski · Charles Beattie · Ellen Clancy · Andrea Huber · Pushmeet Kohli · Andrew Senior · John Jumper · Carl Doersch · S. M. Ali Eslami · Olaf Ronneberger · Jonas Adler -
2021 : Live Panel »
Max Welling · Bharath Ramsundar · Irina Rish · Karianne J Bergen · Pushmeet Kohli -
2021 Poster: Make Sure You're Unsure: A Framework for Verifying Probabilistic Specifications »
Leonard Berrada · Sumanth Dathathri · Krishnamurthy Dvijotham · Robert Stanforth · Rudy Bunel · Jonathan Uesato · Sven Gowal · M. Pawan Kumar -
2020 Poster: Self-Supervised MultiModal Versatile Networks »
Jean-Baptiste Alayrac · Adria Recasens · Rosalia Schneider · Relja Arandjelović · Jason Ramapuram · Jeffrey De Fauw · Lucas Smaira · Sander Dieleman · Andrew Zisserman -
2020 Poster: Training Generative Adversarial Networks by Solving Ordinary Differential Equations »
Chongli Qin · Yan Wu · Jost Tobias Springenberg · Andy Brock · Jeff Donahue · Timothy Lillicrap · Pushmeet Kohli -
2020 Poster: Enabling certification of verification-agnostic networks via memory-efficient semidefinite programming »
Sumanth Dathathri · Krishnamurthy Dvijotham · Alexey Kurakin · Aditi Raghunathan · Jonathan Uesato · Rudy Bunel · Shreya Shankar · Jacob Steinhardt · Ian Goodfellow · Percy Liang · Pushmeet Kohli -
2020 Spotlight: Training Generative Adversarial Networks by Solving Ordinary Differential Equations »
Chongli Qin · Yan Wu · Jost Tobias Springenberg · Andy Brock · Jeff Donahue · Timothy Lillicrap · Pushmeet Kohli -
2020 Poster: The Autoencoding Variational Autoencoder »
Taylan Cemgil · Sumedh Ghaisas · Krishnamurthy Dvijotham · Sven Gowal · Pushmeet Kohli -
2020 Spotlight: The Autoencoding Variational Autoencoder »
Taylan Cemgil · Sumedh Ghaisas · Krishnamurthy Dvijotham · Sven Gowal · Pushmeet Kohli -
2019 Poster: Learning Transferable Graph Exploration »
Hanjun Dai · Yujia Li · Chenglong Wang · Rishabh Singh · Po-Sen Huang · Pushmeet Kohli -
2019 Poster: Adversarial Robustness through Local Linearization »
Chongli Qin · James Martens · Sven Gowal · Dilip Krishnan · Krishnamurthy Dvijotham · Alhussein Fawzi · Soham De · Robert Stanforth · Pushmeet Kohli -
2019 Poster: Learning dynamic polynomial proofs »
Alhussein Fawzi · Mateusz Malinowski · Hamza Fawzi · Omar Fawzi -
2019 Spotlight: Learning dynamic polynomial proofs »
Alhussein Fawzi · Mateusz Malinowski · Hamza Fawzi · Omar Fawzi -
2018 Poster: A Unified View of Piecewise Linear Neural Network Verification »
Rudy Bunel · Ilker Turkaslan · Philip Torr · Pushmeet Kohli · Pawan K Mudigonda -
2018 Poster: Adversarial vulnerability for any classifier »
Alhussein Fawzi · Hamza Fawzi · Omar Fawzi -
2018 Poster: Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding »
Kexin Yi · Jiajun Wu · Chuang Gan · Antonio Torralba · Pushmeet Kohli · Josh Tenenbaum -
2018 Spotlight: Neural-Symbolic VQA: Disentangling Reasoning from Vision and Language Understanding »
Kexin Yi · Jiajun Wu · Chuang Gan · Antonio Torralba · Pushmeet Kohli · Josh Tenenbaum -
2018 Poster: A flexible model for training action localization with varying levels of supervision »
Guilhem Chéron · Jean-Baptiste Alayrac · Ivan Laptev · Cordelia Schmid