DataTerminology.comNEW
A collaborative training approach where a neural network is partitioned across parties, sharing only intermediate activations and gradients.
DESCRIPTION
Split learning partitions a model so that clients train early layers and a server trains later layers, exchanging activations and gradients instead of raw data. This reduces data exposure while enabling centralized training benefits. It suits cross-silo training when data cannot leave sensitive environments. Threats include gradient inversion and reconstruction, mitigated by techniques like noise, encryption, or secure protocols. It complements federated learning but offers different communication and compute trade-offs.
ALTERNATIVE DEFINITIONS
OpenMined
Split learning (SplitNN) is an approach where the neural network is split across multiple parties and only intermediate activations are shared between them during training.
Stresses partitioning and sharing of activations instead of raw data.
arXiv
In split learning, the model is split between the client and server, and the client sends the smashed data (activations) to the server instead of raw data.
Introduces the concept of ‘smashed data’ for privacy.
RELATED TERMS
Privacy-Preserving Machine Learning Client-Server Training Secure Aggregation Differential Privacy Gradient Leakage Edge Computing Model Partitioning Homomorphic Encryption Knowledge Distillation
COMPETING TERMS
KNOWLEDGE AREAS
AI MLOps Data Privacy & Compliance Data Security
TARGET AUDIENCE
AI Engineer MLOps Engineer Data Scientist Chief AI Officer Security Architect
EXTERNAL SOURCES
- Wikipedia - Federated learning#Split learning
- arXiv - Original SplitNN research preprint
- OpenMined - Community tutorial on split learning
- Google - Privacy-preserving ML concepts including variants
TOP QUESTIONS
What is Split Learning and how does it differ from Federated Learning?
Split learning partitions the model, exchanging activations and gradients, while federated learning exchanges model updates. Split can reduce client compute and data exposure differently.
Category: Comparison
What problems does Split Learning solve?
It enables collaborative training when data cannot be centralized due to privacy, regulation, or bandwidth constraints, reducing raw data movement.
Category: Use-case
What are the privacy risks in Split Learning?
Gradient or activation leakage can reveal information. Mitigations include noise addition, secure enclaves, encryption, or differential privacy.
Category: Risks
When should I choose Split Learning over Federated Learning?
Choose split learning when clients are resource-constrained or when server-side layers need centralized control or specialized hardware.
Category: Decision
How do you deploy Split Learning in production?
Use orchestration for client-server training, secure channels, versioned models, and monitor activation/gradient flows and latency.
Category: Operations
What is ‘smashed data’ in Split Learning?
It refers to intermediate activations sent from client to server instead of raw input, intended to minimize information leakage.
Category: Concept
What are performance trade-offs with Split Learning?
It reduces client compute but adds communication overhead and synchronization complexity. Proper cut-layer selection is key.
Category: Limitations
How does Split Learning integrate with MLOps?
Treat client and server segments as coordinated pipelines with versioning, secure artifacts, and monitoring of privacy metrics.
Category: Integration
What is a hot topic around Split Learning?
Combining split learning with differential privacy and secure computation to strengthen privacy while maintaining accuracy.
Category: Topic