Towards Robust Knowledge Removal in Federated Learning with High Data Heterogeneity

TL;DR

The paper tackles the problem of removing a specific client's knowledge from a federated model without significant downtime or retraining. It introduces Stand Alone TA (SATA), a Federated Unlearning approach that combines Task Arithmetic (where client contributions are encoded as task vectors $\tau_t = \theta_t - \theta_0$ and updates follow $\theta_{new} = \theta_0 + \sum_i \lambda_i \tau_i$) with Neural Tangent Kernel (NTK) to linearize updates via $f_{lin}(x; \theta) = f(x; \theta_0) + (\theta - \theta_0)^\top \nabla_\theta f(x; \theta_0)$. A key idea is to train two per-client vectors, a main $\tau_k$ and a standalone $\tau^{sa}_k$, and to unlearn by subtracting $\lambda_{tgt} \tau^{sa}_{tgt}$ from the global model, yielding $\hat{\theta}_{clean} = \hat{\theta} - \lambda_{tgt} \tau^{sa}_{tgt}$, with NTK enabling a linearized, disentangled update $f_{lin}(x; \theta_{r-1} + \sum_k \lambda_k \tau_k - \lambda_{tgt} \tau^{sa}_{tgt})$. The method is validated on Cars-196 and Resisc45 under non-IID settings, showing competitive unlearning performance in a single round while preserving overall accuracy, and is supplemented by a single-round alternative baseline SAFA. Overall, the work demonstrates a practical, fast Federated Unlearning framework with potential broad impact for privacy-compliant, low-disruption FL deployments, and outlines directions for improvement under higher data skew and broader modalities.

Abstract

Nowdays, there are an abundance of portable devices capable of collecting large amounts of data and with decent computational power. This opened the possibility to train AI models in a distributed manner, preserving the participating clients' privacy. However, because of privacy regulations and safety requirements, elimination upon necessity of a client contribution to the model has become mandatory. The cleansing process must satisfy specific efficacy and time requirements. In recent years, research efforts have produced several knowledge removal methods, but these require multiple communication rounds between the data holders and the process coordinator. This can cause the unavailability of an effective model up to the end of the removal process, which can result in a disservice to the system users. In this paper, we introduce an innovative solution based on Task Arithmetic and the Neural Tangent Kernel, to rapidly remove a client's influence from a model.

Figures (1)

• Figure 1: Plots of model accuracy on the Cars196 dataset.