Home > Research > Publications & Outputs > GradAuto

Electronic data

Links

Text available via DOI:

View graph of relations

GradAuto: Energy-oriented Attack on Dynamic Neural Networks

Research output: Contribution in Book/Report/Proceedings - With ISBN/ISSNConference contribution/Paperpeer-review

Published

Standard

GradAuto: Energy-oriented Attack on Dynamic Neural Networks. / Pan, Jianhong; Zheng, Qichen; Fan, Zhipeng et al.
European Conference on Computer Vision (ECCV). ed. / Shai Avidan; Gabriel Brostow; Moustapha Cissé; Giovanni Maria Farinella; Tal Hassner. Springer, 2022. p. 637-653 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 13664).

Research output: Contribution in Book/Report/Proceedings - With ISBN/ISSNConference contribution/Paperpeer-review

Harvard

Pan, J, Zheng, Q, Fan, Z, Rahmani, H, Ke, Q & Liu, J 2022, GradAuto: Energy-oriented Attack on Dynamic Neural Networks. in S Avidan, G Brostow, M Cissé, GM Farinella & T Hassner (eds), European Conference on Computer Vision (ECCV). Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 13664, Springer, pp. 637-653, 17th European Conference on Computer Vision, ECCV 2022, Tel Aviv, Israel, 23/10/22. https://doi.org/10.1007/978-3-031-19772-7_37

APA

Pan, J., Zheng, Q., Fan, Z., Rahmani, H., Ke, Q., & Liu, J. (2022). GradAuto: Energy-oriented Attack on Dynamic Neural Networks. In S. Avidan, G. Brostow, M. Cissé, G. M. Farinella, & T. Hassner (Eds.), European Conference on Computer Vision (ECCV) (pp. 637-653). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 13664). Springer. https://doi.org/10.1007/978-3-031-19772-7_37

Vancouver

Pan J, Zheng Q, Fan Z, Rahmani H, Ke Q, Liu J. GradAuto: Energy-oriented Attack on Dynamic Neural Networks. In Avidan S, Brostow G, Cissé M, Farinella GM, Hassner T, editors, European Conference on Computer Vision (ECCV). Springer. 2022. p. 637-653. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). doi: 10.1007/978-3-031-19772-7_37

Author

Pan, Jianhong ; Zheng, Qichen ; Fan, Zhipeng et al. / GradAuto : Energy-oriented Attack on Dynamic Neural Networks. European Conference on Computer Vision (ECCV). editor / Shai Avidan ; Gabriel Brostow ; Moustapha Cissé ; Giovanni Maria Farinella ; Tal Hassner. Springer, 2022. pp. 637-653 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).

Bibtex

@inproceedings{56a2359fffb3449abc6b263cb7d827a0,
title = "GradAuto: Energy-oriented Attack on Dynamic Neural Networks",
abstract = "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 the computational efficiency without compromising the accuracy. In this paper, we investigate the robustness of dynamic neural networks against energy-oriented attacks. We present a novel algorithm, named GradAuto, to attack both dynamic depth and dynamic width models, where dynamic depth networks reduce redundant computation by skipping some intermediate layers while dynamic width networks adaptively activate a subset of neurons in each layer. Our GradAuto carefully adjusts the direction and the magnitude of the gradients to efficiently find an almost imperceptible perturbation for each input, which will activate more computation units during inference. In this way, GradAuto effectively boosts the computational cost of models with dynamic architectures. Compared to previous energy-oriented attack techniques, GradAuto obtains the state-of-the-art result and recovers 100% dynamic network reduced FLOPs on average for both dynamic depth and dynamic width models. Furthermore, we demonstrate that GradAuto offers us great control over the attacking process and could serve as one of the keys to unlock the potential of the energy-oriented attack. Please visit https://github.com/JianhongPan/GradAuto for code.",
author = "Jianhong Pan and Qichen Zheng and Zhipeng Fan and Hossein Rahmani and Qiuhong Ke and Jun Liu",
year = "2022",
month = oct,
day = "28",
doi = "10.1007/978-3-031-19772-7_37",
language = "English",
isbn = "9783031197710",
series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",
publisher = "Springer",
pages = "637--653",
editor = "Shai Avidan and Gabriel Brostow and Ciss{\'e}, {Moustapha } and Farinella, {Giovanni Maria} and Hassner, {Tal }",
booktitle = "European Conference on Computer Vision (ECCV)",
note = "17th European Conference on Computer Vision, ECCV 2022 ; Conference date: 23-10-2022 Through 27-10-2022",

}

RIS

TY - GEN

T1 - GradAuto

T2 - 17th European Conference on Computer Vision, ECCV 2022

AU - Pan, Jianhong

AU - Zheng, Qichen

AU - Fan, Zhipeng

AU - Rahmani, Hossein

AU - Ke, Qiuhong

AU - Liu, Jun

PY - 2022/10/28

Y1 - 2022/10/28

N2 - 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 the computational efficiency without compromising the accuracy. In this paper, we investigate the robustness of dynamic neural networks against energy-oriented attacks. We present a novel algorithm, named GradAuto, to attack both dynamic depth and dynamic width models, where dynamic depth networks reduce redundant computation by skipping some intermediate layers while dynamic width networks adaptively activate a subset of neurons in each layer. Our GradAuto carefully adjusts the direction and the magnitude of the gradients to efficiently find an almost imperceptible perturbation for each input, which will activate more computation units during inference. In this way, GradAuto effectively boosts the computational cost of models with dynamic architectures. Compared to previous energy-oriented attack techniques, GradAuto obtains the state-of-the-art result and recovers 100% dynamic network reduced FLOPs on average for both dynamic depth and dynamic width models. Furthermore, we demonstrate that GradAuto offers us great control over the attacking process and could serve as one of the keys to unlock the potential of the energy-oriented attack. Please visit https://github.com/JianhongPan/GradAuto for code.

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

U2 - 10.1007/978-3-031-19772-7_37

DO - 10.1007/978-3-031-19772-7_37

M3 - Conference contribution/Paper

SN - 9783031197710

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 637

EP - 653

BT - European Conference on Computer Vision (ECCV)

A2 - Avidan, Shai

A2 - Brostow, Gabriel

A2 - Cissé, Moustapha

A2 - Farinella, Giovanni Maria

A2 - Hassner, Tal

PB - Springer

Y2 - 23 October 2022 through 27 October 2022

ER -