You Only Propagate Once: Painless Adversarial Training Using Maximal Principle
share
Deep learning achieves state-of-the-art results in many areas. However recent works have shown that deep networks can be vulnerable to adversarial perturbations which slightly changes the input but leads to incorrect prediction. Adversarial training is an effective way of improving the robustness to the adversarial examples, typically formulated as a robust optimization problem for network training. To solve it, previous works directly run gradient descent on the "adversarial loss", i.e. replacing the input data with the corresponding adversaries. A major drawback of this approach is the computational overhead of adversary generation, which is much larger than network updating and leads to inconvenience in adversarial defense. To address this issue, we fully exploit structure of deep neural networks and propose a novel strategy to decouple the adversary update with the gradient back propagation. To achieve this goal, we follow the research line considering training deep neural network as an optimal control problem. We formulate the robust optimization as a differential game. This allows us to figure out the necessary conditions for optimality. In this way, we train the neural network via solving the Pontryagin's Maximum Principle (PMP). The adversary is only coupled with the first layer weight in PMP. It inspires us to split the adversary computation from the back propagation gradient computation. As a result, our proposed YOPO (You Only Propagate Once) avoids forward and backward the data too many times in one iteration, and restricts core descent directions computation to the first layer of the network, thus speeding up every iteration significantly. For adversarial example defense, our experiment shows that YOPO can achieve comparable defense accuracy using around 1/5 GPU time of the original projected gradient descent training.
READ FULL TEXT