TABLE X.
Features of Notable TinyML Federated Learning Frameworks
| Framework | FL Strategy | Communication Stack | Scalability and Heterogeneity | Privacy | Client Hardware (language) | Open-source |
|---|---|---|---|---|---|---|
| Flower [190] [191] | FedAvg, Fault tolerant FedAvg, FedProx, QFedAvg, FedAdagrad, FedYogi, FedAdam | Bidirectional gRPC and ClientProxy (language, communication and serialization agnostic) | FedFS (partial work, importance sampling, and dynamic time-outs to handle bandwidth heterogeneity); Virtual Client Engine for scheduling and resource management (15M clients tested) | Salvia secure aggregation | CPU, GPU, MCU (Python, Java, C++) | ✓ |
| FedPARL [192] | Reparametrized FedAvg with sample-based pruning | None (simulated) | Resource tracking (memory, battery life, bandwidth, and data volume); Trust value tracking (task completion, delay, model integrity); Partial work (12 clients tested) | Vanilla model aggregation | None (simulated) | χ |
| DIoT [193] | FedAvg | Bidirectional WebSocket protocol over WiFi and Ethernet | AuDI device-type identification (15 clients tested) | Vanilla model aggregation | CPU, GPU (Python and JavaScript) | χ |
| PruneFL [194] | FedAvg with adaptive and distributed pruning | WiFi and Ethernet, with distributed pruning to reduce communication overhead | Adaptive pruning to modify local models based on resource availability (9 clients tested) | Vanilla model aggregation | CPU, MCU (Python) | ✓ |
| TinyFedTL [195] | FedAvg with last layer transfer learning | USART | 9 clients tested | Vanilla model aggregation | MCU (C++) | ✓ |
| FLAgr [196] | Reinforcement learning | None (simulated) | Real-time collaboration scheme discovery via deep deterministic policy gradient (1000 clients tested) | Rating feedback mechanism | None (simulated) | χ |
| PerFit [197] | FedPer, FedHealth, FedAvg, Personalized FedAvg, MOCHA, FedMD, Federated Distillation | WiFi, BLE, Cellular (simulated) | Federated transfer learning, federated distillation, federated meta-learning, and federated multi-task learning to personalize the model, device and statistical heterogeneity (30 clients tested) | Vanilla model aggregation | None (simulated) | χ |