[5N2-GS-11h] AI and Society

1:15 PM - 1:30 PM JST(4:15 AM - 4:30 AM UTC)

[5N2-GS-11h-06] FedC-CHVAE: Generating Plausible Counterfactual Explanations in Federated Settings

〇Akihiro Toyoda¹, Yuji Kawamata^2,3, Yukihiko Okada^2,3 (1. Graduate School of Science and Technology, Univ. of Tsukuba, 2. Institute of Systems and Information Engineering, Univ. of Tsukuba, 3. Center for Artificial Intelligence Research, Tsukuba Institute for Advanced Research, Univ. of Tsukuba)

Keywords:

federated learning,counterfactual explanation,explainability

In highly regulated domains such as finance and healthcare, counterfactual explanations (CE) can improve transparency by suggesting actionable changes that overturn an unfavorable prediction. However, in federated learning (FL) settings where data remain distributed across organizations, generating plausible counterfactuals is challenging because centralizing queries or generated candidates can leak sensitive information, and each client alone may lack sufficient data to learn the data distribution required by generative CE methods. We propose FedC-CHVAE, a federated extension of C-CHVAE. Each client trains a VAE conditioned on immutable features, while the server aggregates only weights via FedAvg. Counterfactual search and generation are performed locally on each client. Experiments on the Adult dataset with three IID clients show that FedC-CHVAE improves proximity, sparsity, and manifold distance over local training and approaches centralized training performance. These results demonstrate the practicality of privacy-preserving, distribution-faithful CEs in federated settings.