Cross-Device Federated Learning

Since raw data never leaves the device, privacy is architectural. However, model updates can still leak information, requiring additional safeguards.

• Differential Privacy (DP): Noise is added to local updates before transmission to provide plausible deniability.
• Secure Aggregation: A cryptographic protocol ensures the server can only compute the sum of updates, never inspecting individual contributions.
• Data Locality: All training data—keystrokes, health metrics, voice snippets—remains on-device permanently.

The contrasting topology where a small number of reliable, stateful institutional clients (e.g., hospitals, banks) collaboratively train a model. Unlike cross-device, silos hold large, curated datasets and are almost always available. Key distinctions include:

• Scale: Tens to hundreds of clients vs. millions.
• Reliability: High availability vs. frequent client dropout.
• Compute: Powerful servers vs. resource-constrained smartphones.

The strategic algorithm that determines which subset of thousands of available edge devices participates in a given training round. Effective selection must balance:

• Device Capability: Filtering out devices with low battery or metered connections.
• Statistical Representativeness: Ensuring the selected cohort reflects the overall data distribution to avoid bias.
• Fairness: Preventing over-reliance on a small pool of power users to ensure the global model generalizes broadly.

A communication protocol where the central server updates the global model immediately upon receiving an update from any single device, without waiting for a full cohort. This is critical for cross-device topologies because:

• It eliminates the straggler bottleneck caused by slow or intermittently connected devices.
• It enables continuous model improvement from a firehose of heterogeneous, unpredictable edge updates.
• Requires sophisticated staleness mitigation to prevent outdated updates from degrading the global model.

A cryptographic protocol ensuring the central server can compute the sum of model updates without inspecting any individual contribution. In cross-device settings with millions of untrusted nodes, this is non-negotiable. The protocol uses secret sharing where devices mask their updates with pairwise noise that cancels out during summation, guaranteeing that even a compromised server cannot isolate a single user's gradient vector.

Techniques to prevent slow or unresponsive edge devices from blocking the entire synchronous training round. Common strategies include:

• Timeouts: Discarding updates from devices that exceed a predefined deadline.
• Backup Workers: Over-selecting clients and accepting the first k responses.
• Asynchronous Fallback: Dynamically switching to an async protocol when a threshold of stragglers is detected, ensuring wall-clock training time remains predictable.

The application of gradient compression, quantization, and sparsification to reduce the communication payload from edge devices. Since cross-device clients often upload over cellular networks, minimizing bits per round is essential. Techniques include:

• Gradient Sparsification: Transmitting only the top-k largest gradient values.
• Probabilistic Quantization: Reducing 32-bit floats to 2-bit or 1-bit representations.
• Error Feedback: Retaining compression residuals locally to prevent accuracy degradation over successive rounds.