Evaluation of Data Poisoning Attack on Centralized and Federated Learning environments.

Safiia Mohammed

Keywords: Data poisoning attack, Federated Learning, Flower FL, Centralized Machine Learning, Machine Learning.

Abstract:

This paper investigates the effectiveness of data poisoning attacks in centralized
and federated learning environments. The research utilizes the Flower framework
to establish a federated learning setting, which introduces unique challenges and
possibilities for malicious actors.

The evaluation involves comparing the impact of data poisoning attacks on two
datasets, CIFAR10 and MNIST—the attack success rate used as a metric to evaluate
the efficacy of the attacks in both environments. The results indicate that federated
learning exhibits higher resistance to data poisoning attacks when applied to the
CIFAR10 dataset. However, centralized learning shows a slightly higher resilience
level than federated learning when applied to the MNIST dataset.

Submission Category: Machine learning algorithms

Submission Number: 49