Federated Learning Data Deletion via Krylov Subspace Influence Approximation
A new method called HF-KCU enables efficient data deletion in federated learning systems by approximating the influence function using conjugate gradient iterations in Krylov subspaces. This reduces computational complexity from O(d³) to O(kd) where k is much smaller than d, making it feasible to remove a client's contribution without retraining from scratch. The approach addresses privacy regulation compliance for data deletion requests in collaborative optimization settings, including adversarial contributions.
Key facts
- HF-KCU removes a client's contribution by approximating the influence function
- Uses conjugate gradient iterations in Krylov subspaces
- Reduces complexity from O(d³) to O(kd)
- k is much smaller than d
- Addresses data deletion requests in federated learning
- Complies with privacy regulations
- Avoids retraining from scratch
- Handles adversarial contributions
Entities
Institutions
- arXiv