New adversarial attacks target federated GANs with label flipping

Researchers have developed new adversarial attacks targeting federated learning setups for Generative Adversarial Networks (GANs). These attacks, including label flipping and an oversampling variant, aim to manipulate the global generator by altering label information during local training. The goal is to skew the learned distribution, causing samples conditioned on a target label to be mapped to a source class. Analysis shows that the semantic damage of these attacks grows linearly with poisoning strength, while deviations from the true target distribution increase quadratically, making them effective yet difficult to detect. AI

IMPACT These attacks highlight potential vulnerabilities in federated learning for generative models, necessitating improved defenses against data poisoning and manipulation.

RANK_REASON The cluster contains a research paper detailing new adversarial attacks on federated GANs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• CIFAR-10
• federated learning
• FEMNIST
• Gans
• Kullback--Leibler divergence
• label flipping
• MNIST database
• oversampling

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →