GradAuto: Energy-Oriented Attack on Dynamic Neural Networks

Dynamic neural networks could adapt their structures or parameters based on different inputs. By reducing the computation redundancy for certain samples, it can greatly improve computational efficiency without compromising accuracy. In this paper, we investigate robustness of dynamic against energy-oriented attacks. We present a novel algorithm, named GradAuto, to attack both depth and width models, where reduce redundant by skipping some intermediate layers while adaptively activate subset neurons in each layer. Our GradAuto carefully adjusts direction magnitude gradients efficiently find an almost imperceptible perturbation input, which will more units during inference. way, effectively boosts cost models with architectures. Compared previous techniques, obtains state-of-the-art result recovers 100% network reduced FLOPs average models. Furthermore, demonstrate that offers us great control over attacking process serve as one keys unlock potential attack. Please visit https://github.com/JianhongPan/GradAuto code.