arXiv (Cornell University) · 2017 · 433 citations · 19 references
Open access
ComputingComputer ScienceComputer SecurityComputer VisionMachine LearningConvex OptimizationDeep LearningLoss FunctionNeural NetworksOptimization For Machine LearningRepresentation LearningAdversarial Detection RobustnessAdversarial ExamplesAdversarial LossAdversarial RobustnessBackpropagationRobust Convex OptimizationRobust LossRobust Loss Functions
The authors propose a method to train deep ReLU classifiers that are provably robust to norm‑bounded adversarial perturbations during training. They achieve this by formulating a convex outer approximation of the activation set under norm‑bounded perturbations and solving a linear‑program‑based robust optimization whose dual can be implemented as a modified back‑propagation network, enabling efficient training with only a few extra passes. The resulting classifiers detect all adversarial examples on unseen data (with some false positives) and achieve provably low test error—e.g., less than 5.8 % on MNIST for ε = 0.1—while the training procedure requires only a modest increase in computation.
We propose a method to learn deep ReLU-based classifiers that are provably robust against norm-bounded adversarial perturbations on the training data. For previously unseen examples, the approach is guaranteed to detect all adversarial examples, though it may flag some non-adversarial examples as well. The basic idea is to consider a convex outer approximation of the set of activations reachable through a norm-bounded perturbation, and we develop a robust optimization procedure that minimizes the worst case loss over this outer region (via a linear program). Crucially, we show that the dual problem to this linear program can be represented itself as a deep network similar to the backpropagation network, leading to very efficient optimization approaches that produce guaranteed bounds on the robust loss. The end result is that by executing a few more forward and backward passes through a slightly modified version of the original network (though possibly with much larger batch sizes), we can learn a classifier that is provably robust to any norm-bounded adversarial attack. We illustrate the approach on a number of tasks to train classifiers with robust adversarial guarantees (e.g. for MNIST, we produce a convolutional classifier that provably has less than 5.8% test error for any adversarial attack with bounded $\ell_\infty$ norm less than $ε= 0.1$), and code for all experiments in the paper is available at https://github.com/locuslab/convex_adversarial.
19
Adversarial Examples in the Physical World
Alexey Kurakin, Ian Goodfellow, Samy Bengio · 2018 · 1.9K citations · Full text
Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer et al. · 2016 · 1.6K citations · Full text