Game of Threads: Enabling Asynchronous Poisoning Attacks

Published in ASPLOS, 2020

Full Paper | bibtex | Plain Text

@inproceedings{10.1145/3373376.3378462,
author = {Sanchez Vicarte, Jose Rodrigo and Schreiber, Benjamin and Paccagnella, Riccardo and Fletcher, Christopher W.},
title = {Game of Threads: Enabling Asynchronous Poisoning Attacks},
year = {2020},
isbn = {9781450371025},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
url = {https://doi.org/10.1145/3373376.3378462},
doi = {10.1145/3373376.3378462},
booktitle = {Proceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems},
pages = {35–52},
numpages = {18},
keywords = {adversarial machine learning, trusted execution environment, asynchronous stochastic gradient descent},
location = {Lausanne, Switzerland},
series = {ASPLOS ’20}
}
Jose Rodrigo Sanchez Vicarte, Benjamin Schreiber, Riccardo Paccagnella, and Christopher W. Fletcher. 2020. Game of Threads: Enabling Asynchronous Poisoning Attacks. In Proceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS ’20). Association for Computing Machinery, New York, NY, USA, 35–52. DOI:https://doi.org/10.1145/3373376.3378462
Coming soon
PyTorchFI: A Runtime Perturbation Tool for DNNs; Abdulrahman Mahmoud, Neeraj Aggarwal, Alex Nobbe, Jose Rodrigo Sanchez Vicarte, Sarita Adve, Christopher W. Fletcher, Iuri Frosio, Siva Kumar Sastry Hari; DSN-DSML, 2020
@inproceedings{10.1109/ISCA.2018.00053,
author = {Choi, Woo-Seok and Tomei, Matthew and Vicarte, Jose Rodrigo Sanchez and Hanumolu, Pavan Kumar and Kumar, Rakesh},
title = {Guaranteeing Local Differential Privacy on Ultra-Low-Power Systems},
year = {2018},
isbn = {9781538659847},
publisher = {IEEE Press},
url = {https://doi.org/10.1109/ISCA.2018.00053},
doi = {10.1109/ISCA.2018.00053},
booktitle = {Proceedings of the 45th Annual International Symposium on Computer Architecture},
pages = {561–574},
numpages = {14},
keywords = {microcontrollers, randomized response, IoT, low-power systems, RAPPOR, differential privacy},
location = {Los Angeles, California},
series = {ISCA ’18}
}
Woo-Seok Choi, Matthew Tomei, Jose Rodrigo Sanchez Vicarte, Pavan Kumar Hanumolu, and Rakesh Kumar. 2018. Guaranteeing local differential privacy on ultra-low-power systems. In Proceedings of the 45th Annual International Symposium on Computer Architecture (ISCA ’18). IEEE Press, 561–574. DOI:https://doi.org/10.1109/ISCA.2018.00053

artifact available badge artifact evaluated and functional badge

As data sizes continue to grow at an unprecedented rate, machine learning training is being forced to adopt asynchronous algorithms to maintain performance and scalability. In asynchronous training, many threads share and update the model in a racy fashion to avoid costly inter-thread synchronization.

This paper studies the security implications of these codes by introducing asynchronous poisoning attacks. Our attack influences training outcome—e.g., degrades model accuracy or biases the model towards an adversary-specified label—purely by scheduling asynchronous training threads in a malicious fashion. Since thread scheduling is outside the protections of modern trusted execution environments (TEEs), e.g., Intel SGX, our attack bypasses these protections even when the training set can be verified as correct. To the best of our knowledge, this represents the first example where a class of applications loses integrity guarantees, despite being protected by enclave-based TEEs such as SGX.

We demonstrate both accuracy degradation and model biasing attacks on the CIFAR-10 image recognition task, trained on Resnet-style DNNs using an asynchronous training code published by Pytorch. We also perform proof-of-concept experiments to validate our assumptions on an SGX-enabled machine. Our accuracy degradation attacks are capable of returning a converged model to pre-trained accuracy or to some accuracy in between. Our model biasing attack can force the model to predict an adversary-specified label up to ~40% of the time on the CIFAR-10 validation set (whereas the un-attacked model’s prediction rate towards any label is ~10%).

Artifact available here

github repository available here