Evaluating the robustness of a defense model is a challenging task in adversarial robustness research. Obfuscated gradients have previously been found to exist in many defense methods and cause a false signal of robustness. In this paper, we identify a more subtle situation called Imbalanced Gradients that can also cause overestimated adversarial robustness. The phenomenon of imbalanced gradients occurs when the gradient of one term of the margin loss dominates and pushes the attack towards to a suboptimal direction. To exploit imbalanced gradients, we formulate a Margin Decomposition (MD) attack that decomposes a margin loss into individual terms and then explores the attackability of these terms separately via a two-stage process. We also propose a multi-targeted and ensemble version of our MD attack. By investigating 24 defense models proposed since 2018, we find that 11 models are susceptible to a certain degree of imbalanced gradients and our MD attack can decrease their robustness evaluated by the best standalone baseline attack by more than 1%. We also provide an in-depth investigation on the likely causes of imbalanced gradients and effective countermeasures.

Deep neural networks (DNNs) are known for their vulnerability to adversarial examples. These are examples that have undergone a small, carefully crafted perturbation, and which can easily fool a DNN into making misclassifications at test time. Thus far, the field of adversarial research has mainly focused on image models, under either a white-box setting, where an adversary has full access to model parameters, or a black-box setting where an adversary can only query the target model for probabilities or labels. Whilst several white-box attacks have been proposed for video models, black-box video attacks are still unexplored. To close this gap, we propose the first black-box video attack framework, called V-BAD. V-BAD is a general framework for adversarial gradient estimation and rectification, based on Natural Evolution Strategies (NES). In particular, V-BAD utilizes \textit{tentative perturbations} transferred from image models, and \textit{partition-based rectifications} found by the NES on partitions (patches) of tentative perturbations, to obtain good adversarial gradient estimates with fewer queries to the target model. V-BAD is equivalent to estimating the projection of an adversarial gradient on a selected subspace. Using three benchmark video datasets, we demonstrate that V-BAD can craft both untargeted and targeted attacks to fool two state-of-the-art deep video recognition models. For the targeted attack, it achieves $>$93\% success rate using only an average of $3.4 \sim 8.4 \times 10^4$ queries, a similar number of queries to state-of-the-art black-box image attacks. This is despite the fact that videos often have two orders of magnitude higher dimensionality than static images. We believe that V-BAD is a promising new tool to evaluate and improve the robustness of video recognition models to black-box adversarial attacks.