On-Device Training

All training data remains resident on the edge device's persistent storage or volatile memory. Raw sensor data, keystrokes, images, and audio never leave the device, ensuring compliance with data residency regulations. The model ingests local data streams directly, performing forward and backward passes on-device. This eliminates network transfer latency and removes the attack surface associated with centralized data lakes. Stateless aggregation is the only permissible egress, typically in the form of encrypted gradient vectors or weight deltas.

Edge hardware operates under severe compute, memory, and energy budgets compared to cloud infrastructure. On-device training requires:

• Peak memory footprint often limited to <100MB for mobile SoCs
• Compute budgets measured in megaflops, not teraflops
• Thermal envelopes that prevent sustained high utilization
• Battery impact must remain imperceptible to the user Frameworks like TensorFlow Lite and CoreML enforce these constraints through operator-level profiling and selective layer freezing.

Training on edge devices is inherently opportunistic and interruptible. Unlike server-side training with guaranteed uptime, on-device optimization must contend with:

• Intermittent connectivity requiring graceful pause and resume
• Device idle states as the primary training window (charging, screen-off)
• Operating system preemption that can kill processes at any time Checkpointing and state serialization are mandatory to prevent loss of partial progress. Training is designed as a background daemon, not a foreground critical path.

A globally pre-trained base model is shipped to the device and subsequently adapted to the individual user's patterns. This enables:

• Next-word prediction that learns personal vocabulary and slang
• Face recognition that adapts to changes in appearance over time
• Activity classification tuned to an individual's gait or movement signature Techniques like Parameter-Efficient Fine-Tuning (PEFT) with low-rank adaptation (LoRA) update only a tiny fraction of weights, preserving the base model's generalization while enabling rapid personalization.

Even though raw data stays local, shared model updates can leak information about the training set through gradient inversion attacks. On-device training pipelines integrate:

• Differential privacy via per-sample gradient clipping and Gaussian noise injection
• Secure aggregation protocols that encrypt updates before transmission
• Local differential privacy (LDP) where noise is applied on-device before any data leaves These guarantees are formalized through a privacy budget (epsilon) that quantifies the maximum information leakage per training round.

On-device training is rarely a standalone process; it is the client-side component of a federated learning system. The device participates in:

• Client selection rounds where the server chooses which devices contribute
• Local epoch execution with a configurable number of SGD steps
• Model delta uploads containing only the difference between the initial and updated weights
• Global model synchronization where the device periodically receives the latest aggregated model The server handles straggler mitigation, dropping devices that fail to report updates within a defined time window.