Our attack is carried out in two distinct phases: backdoor injection and backdoor stealthiness. Experimental results using ResNet-18 and MobileNet-V2 models trained on CIFAR-10 and Tiny ImageNet datasets show the effectiveness of our proposed attack. The results highlight the success of our attack in increasing energy consumption on trigger samples while preserving the default model’s performance for clean/regular inputs.
-
Publication
« Energy backdoor attack to deep neural networks »
-
Article
The transition to deep learning has led to a significant surge in computing complexity and energy consumption. This trend has prompted the adoption of acceleration hardware like tensor processing units (TPUs) and application specific integrated circuits (ASICs) to facilitate energy-efficient deployment on edge and mobile devices. However, recent studies have revealed their vulnerability to energy attacks. In this paper, we design an innovative energy backdoor attack against deep neural networks operating on sparsity-based accelerators.