.png){kind=icon}
.png){kind=icon}
Return
SAFE-FL: Secure and Adaptive Federated Encryption for Energy-Efficient Learning
.png){kind=icon}
OA
AI
Z
A
Y
R
D
DOI：10.1109/tgcn.2026.3711489.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Battery-powered IoT devices provide rich local data but operate under strict energy and privacy constraints. Federated learning (FL) keeps raw data on-device, while CKKS homomorphic encryption enables encrypted aggregation of model updates; however, CKKS can significantly increase computation, ciphertext size, latency, and energy consumption on constrained nodes. We propose SAFE-FL, an orchestration framework for encrypted federated learning under battery constraints that jointly adapts CKKS parameters, battery-aware client selection, and per-round duty-cycle duration under battery availability, training feasibility, and 128-bit security constraints. We formulate the design as a mixed-integer multi-objective optimization problem and derive a closed-form control policy for online operation. On MNIST, SAFE-FL reaches a target accuracy 38% faster and preserves 11% more residual battery energy than a fixed-duty-cycle, fixed-CKKS baseline, and up to 67% more energy than an adaptive duty-cycle scheme without CKKS retuning at the same security level. SAFE-FL achieves accuracy comparable to FedSHE (CKKS-based FL with segmented HE) and consistently higher than FedPHE (packed-CKKS HE-enabled FL) on MNIST and CIFAR-10, while improving resilience under intermittent energy availability. The policy evaluates in under 0.4 ms—about seven orders of magnitude faster than exhaustive search—enabling practical encrypted FL for resource-constrained IoT devices.
Keywords:
Federated learning
homomorphic encryption
CKKS
IoT
energy-aware optimization
duty cycling
client selection
Journal
I
IF:
6.7
Papers:
1.3K
Citations:
4.3K
Organization
Cited Papers