Skip to content

Latest commit

 

History

History
53 lines (44 loc) · 1.6 KB

README.md

File metadata and controls

53 lines (44 loc) · 1.6 KB

Introduction

byzantine_resilience module inclues two files: defense.py and attack.py

defense.py includes some aggregate algorithms:

  • Avg
  • Trimmed-mean
  • Median
  • Krum
  • MKrum
  • Bulyan
  • Dnc

attack.py includes Gaussian and ARG-TAL algorithms to get malicious gradients:

  • AGR-TAL
  • Gaussian

Usage

In Byzantine_defense.sh, add follow parameters:

  • --use_gradient
    client send gradients to server
  • --byzantine_aggregate
    use byzantine defense module
  • --by_attack Gaussian
    use Gaussian attack method, or ARG-TAL
  • --by_defense Krum
    use Krum defense method, or Avg, Trimmed-mean, Median, MKrum, Bulyan, Dnc
  • --num_worker 10
    total clients number
  • --by_workers 2
    byzantine clients number
  • --all_grads
    all_grads means byzantine clients know all gradients of clients, default False

References

[1] Xie C, Koyejo O, Gupta I. Phocas: dimensional byzantine-resilient stochastic gradient descent[J]. arXiv preprint arXiv:1805.09682, 2018.

[2] Blanchard P, El Mhamdi E.M., et al. Machine learning with adversaries: Byzantine tolerant gradient descent[C]. International Conference on Neural Information Processing Systems,2017:118-128.

[3] Mhamdi E M E, Guerraoui R, Rouault S. The hidden vulnerability of distributed learning in Byzantium[C]. International Conference on Machine Learning. 2018 : 3521–3530.

[4] Shejwalkar, V. and Houmansadr, A. Manipulating the Byzantine: Optimizing Model Poisoning Attacks and Defenses for Federated Learning[C]. The Network and Distributed System Security Symposium. Internet Society,2021.