Time-Constrained Adversarial Defense in IoT Edge Devices through Kernel Tensor Decomposition and Multi-DNN Scheduling

The development of deep learning technology has resulted in great contributions in many artificial intelligence services, but adversarial attack techniques on deep learning models are also becoming more diverse and sophisticated. IoT edge devices take cloud-independent on-device DNN (deep neural network) processing technology to exhibit a fast response time. However, if the computational complexity of the denoizer for adversarial noises is high, or if a single embedded GPU is shared by multiple DNN models, adversarial defense at the on-device level is bound to represent a long latency. To solve this problem, eDenoizer is proposed in this paper. First, it applies Tucker decomposition to reduce the computational amount required for convolutional kernel tensors in the denoizer. Second, eDenoizer effectively orchestrates both the denoizer and the model defended by the denoizer simultaneously. In addition, the priority of the CPU side can be projected onto the GPU which is completely priority-agnostic, so that the delay can be minimized when the denoizer and the defense target model are assigned a high priority. As a result of confirming through extensive experiments, the reduction of classification accuracy was very marginal, up to 1.78%, and the inference speed accompanied by adversarial defense was improved up to 51.72%.