Definition
Federated Learning (FL) is a decentralized machine learning approach in which training occurs directly on local devices - such as edge nodes, cameras, medical systems, or on-premise servers - without transferring raw data to a central location. Only model updates (weights or gradients) are sent to a coordinating server, which aggregates them into a global model.
This architecture enables training on sensitive or regulated datasets that cannot be shared between organizations, regions, or devices due to legal, privacy, or operational constraints.
Architecture and workflow
- Global model initialization - a server distributes a base model to participating clients.
- Local training - each client trains the model on its local dataset (e.g., video frames, images, metadata).
- Upload of updates - only gradients or parameters are sent back to the server.
- Aggregation - the server computes a new global model version (commonly using FedAvg).
- Re-distribution - the updated global model is sent back to clients.
Key standards and reference points
- FedAvg algorithm - seminal work by McMahan et al. (2017).
- Frameworks such as OpenFL (Intel) and TensorFlow Federated.
- Privacy-preserving ML guidelines combining FL with Differential Privacy or Secure Multi-Party Computation.
Technical metrics and evaluation parameters
Metric | Meaning |
Divergence | Difference between global and local models; affects stability of training. |
Communication Overhead | Amount of data exchanged per training round. |
Round Latency | Total time required for one global update cycle. |
Non-IID variability | Degree to which client data distributions differ. |
Advantages
- Strong privacy guarantees - raw data never leaves the device.
- Regulatory compliance - supports GDPR, HIPAA, and similar frameworks.
- Scalable to distributed, heterogeneous systems - suitable for fleets of cameras or sensors.
- Reduced bandwidth usage - only model updates are transmitted.
Challenges and limitations
- Non-IID distributions often reduce accuracy or slow convergence.
- Potential gradient leakage attacks enabling inference of personal data.
- Variable compute capabilities on heterogeneous devices.
- High communication frequency for large models.
Relevance in image and video anonymization
Federated Learning is highly relevant for privacy-sensitive visual systems because it enables training of detection and anonymization models without centralizing raw video data. Key advantages in this context include:
- local training of face, license plate, or object detectors on device-captured video,
- constant model improvement without exporting identifiable video frames,
- privacy-by-design workflows ensuring data minimization,
- adaptation of models to local conditions (lighting, camera angle, scene dynamics).
Typical applications include:
- smart-city surveillance systems updating models on-device,
- medical imaging anonymization workflows where frames cannot be exported,
- fleet vehicle camera systems learning to detect plates under varying conditions,
- industrial video analytics where strict confidentiality prevents data sharing.