Reproduce “PCDM: A Diffusion-Based Data Poisoning Attack Against Federated Learning Systems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wei Sun, Yijun Chen, Bo Gao, Ke Xiong, Yuwei Wang, Pingyi Fan, Khaled Ben Letaief
PCDM introduces a novel diffusion-based data poisoning attack framework for Federated Learning (FL) that generates highly effective yet stealthy poisoned data by integrating a poisoning vector into a conditional diffusion model and employing a jumping diffusion strategy. The attack consistently degrades global model accuracy across various FL settings and datasets, including real-world wireless benchmarks, while successfully evading eleven state-of-the-art defense mechanisms and achieving an 85% reduction in computational latency compared to standard diffusion models.
Key insights
- Diffusion models, due to their stochastic denoising process and fine-grained control, can overcome the mode collapse limitations of GANs, generating poisoned data that is statistically indistinguishable from benign data.
- The poisoning vector (v) acts as a critical component, systematically improving both the stealthiness and aggressiveness of the attack by introducing subtle, controlled adversarial gradients during the reverse diffusion process.
- Optimizing the diffusion process through a 'jumping diffusion strategy' allows for significant computational efficiency gains without sacrificing attack potency, making advanced generative attacks viable for resource-limited FL clients.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.