Fed-FBD: New federated learning blocks enable surgical unlearning in seconds

Federated learning has long struggled with three core problems: isolating adversarial contributors, auditing influence, and honoring the right to be forgotten. Standard approaches like FedAvg treat each client as a black box, leaving no mechanism for surgical removal. To address this, researchers Weijie Chen and Alan B. McMillan present Fed-FBD (Federated Functional Block Diversification), a modular architecture that decomposes a ResNet backbone into six functional blocks—the stem, four residual groups, and the classification head. Instead of merging all client updates into a single global model, Fed-FBD maintains a warehouse of N 'color variants,' each assembled from independently tracked and contributor-stamped blocks. This design automatically confines any adversarial or mislabeled client's influence to its own blocks, preventing contamination of clean colors.

On the privacy front, Fed-FBD provides built-in protection: membership inference advantage is already indistinguishable from chance before any formal privacy mechanism is applied. For unlearning, the architecture allows the complete removal of a departed participant's contribution at sub-second cost—simply by deleting that client's block contributions—without any retraining. Experiments across six MedMNIST-2D datasets, PathMNIST at 224x224, and CIFAR-10 show the trade-offs are modest: a 0.3-3.1% IID accuracy gap on adequately sized datasets, and within 0.8-4.0% of FedAvg under non-IID Dirichlet alpha=1.0 on three of four datasets. All six tested adversarial attacks remained confined to the poisoned client's blocks, with at most +/-0.01 AUC drift on clean colors.