Work

FedUP: Efficient pruning-based federated unlearning for model poisoning attacks

Year: 2026
Type: article
Abstract: Federated Learning (FL) can be vulnerable to attacks such as model poisoning, where adversaries send malicious local weights to compromise the global model. Federated Unlearning (FU) is emerging as a solution to address such vulnerabilities by selectively removing the influence of detected malicious contributors without complete retraining. However, unlike typical FU scenarios where clients are trusted and cooperative, applying FU with malicious and possibly colluding clients is challenging because their collaboration in unlearning cannot be assumed. This work presents FedUP, a lightweight FU algorithm designed to efficiently mitigate malicious clients’ influence by pruning specific connections within the attacked model. FedUP achieves efficiency by relying only on clients’ weights from the last training round before unlearning to identify which connections to inhibit. Isolating malicious influence is non-trivial because benign and malicious updates may overlap. FedUP addresses this by selecting and zeroing the highest-magnitude weights that diverge most between the latest updates from benign and malicious clients while preserving benign information. We evaluate FedUP under a strong adversarial threat model, where up to 50 % − 1 of the clients may be malicious and have full knowledge of the aggregation process, including adaptive attackers aware of FedUP’s pruning strategy. The evaluation covers IID and Non-IID data, label-flipping and backdoor attacks, imperfect detection, and comparisons with state-of-the-art FU solutions. In our experiments, FedUP reduces malicious influence to levels comparable to retraining from scratch while preserving benign performance, achieving effective unlearning while consistently being faster and requiring less storage than the state of the art.
Language: English
Open Access status: hybrid
Estimated APC paid: $3,130
OpenAlex ID: W4415013432
FWCI: 3.769
Cites: 75
Cited by: 5
Related to: 10