Personalized Federated Learning

The most straightforward personalization strategy where a globally trained model serves as an initialization for local adaptation. After standard Federated Averaging (FedAvg) converges, each client performs additional training epochs on its private Non-IID Data. This bridges the gap between the general global representation and the client's specific distribution without requiring the server to manage multiple model states. The technique is computationally lightweight but offers no formal protection against catastrophic forgetting of global knowledge.

This approach treats each client as a distinct task in a Multi-Task Learning framework, learning shared representations while allowing for client-specific divergence. Instead of enforcing a single consensus, the objective function explicitly models the relationships between local models using regularization. Techniques like MOCHA (Model-Agnostic Clustered Federated Learning) exploit task relatedness, grouping clients with similar data distributions to learn robust cluster-specific models, effectively handling Concept Drift across the network.

A parameter-efficient personalization method that creates a personalized model by interpolating between a global model and a local model. The mixing weight, often learned dynamically, balances global generalization against local specialization. APFL (Adaptive Personalized Federated Learning) implements this by optimizing a convex combination: θ_personal = α * θ_local + (1 - α) * θ_global. This prevents the local model from overfitting to a small dataset while capturing individual nuances.

Instead of personalizing the entire model, this strategy designates specific layers as personal and keeps them local, while sharing the remaining layers via Federated Averaging. Typically, base layers (closer to the input) learn general feature extractors and are shared, while head layers (classifier) are kept private to learn user-specific mappings. FedPer is a canonical example, reducing communication costs and explicitly preventing the global model from diluting personalized decision boundaries.

Leverages Model-Agnostic Meta-Learning (MAML) to find a global initialization that is highly sensitive to local data. The server trains a meta-model that can rapidly adapt to a new client's distribution with only one or a few gradient steps. Per-FedAvg modifies the FedAvg objective to approximate the MAML gradient, explicitly optimizing for personalization post-aggregation. This is ideal for cold-start problems where new clients join the network and need immediate high performance.

Uses Federated Distillation to exchange knowledge without sharing model weights. Clients share soft predictions (logits) on a public, unlabeled dataset instead of gradients. A local model is trained to mimic the ensemble of global predictions while fitting private data. This is structurally immune to Gradient Inversion attacks and naturally handles heterogeneous model architectures, allowing different clients to use completely different neural network designs tailored to their hardware constraints.

The primary catalyst for personalization. In federated learning, local datasets are Non-Independently and Identically Distributed, meaning the data on one client (e.g., a user's photo library) does not statistically represent the global population.

• Label skew: One client has mostly cats, another mostly dogs.
• Feature skew: Handwriting varies drastically between users.
• Concept drift: User behavior changes over time. Personalized FL addresses this by moving beyond a single global model to fit local distributions.

The de facto baseline algorithm for federated optimization. Clients train locally on their private data and send only model weight updates to the server, which computes a weighted average to form a new global model.

• Minimizes communication rounds.
• Assumes a single global model fits all clients.
• Performance degrades severely under Non-IID data. Personalized FL variants often use FedAvg as a starting point before applying local fine-tuning or multi-task learning to correct for client drift.

A meta-learning algorithm designed to find a model initialization that can adapt rapidly to new tasks with only a few gradient steps. In personalized FL, MAML trains a global initialization such that a single step of local SGD produces a high-performance personalized model.

• Inner loop: Client adapts to local data.
• Outer loop: Server optimizes for fast adaptability. This approach explicitly optimizes for personalization rather than treating it as an afterthought, making it robust to statistical heterogeneity.

A framework that treats each client as a distinct but related task. Instead of forcing a single consensus model, multi-task personalized FL learns a shared representation while allowing client-specific model parameters to diverge.

• MOCHA: A communication-efficient algorithm for multi-task federated learning.
• Handles statistical heterogeneity by design.
• Exploits task relationships to improve generalization for clients with limited data. This is a principled alternative to simple local fine-tuning, providing theoretical guarantees on convergence.

A cryptographic protocol that ensures the central server can only compute the sum of client model updates without ever inspecting individual contributions. This is critical for personalized FL because local models encode highly sensitive user-specific patterns.

• Uses Secret Sharing and Key Agreement.
• Masks individual updates with pairwise random masks that cancel out in the sum.
• Protects against a curious-but-honest server. Without secure aggregation, gradient leakage attacks could reconstruct private training data from personalized model updates.

A mathematical framework providing provable privacy guarantees by injecting calibrated noise into model updates. In personalized FL, local differential privacy ensures that a client's fine-tuned model does not memorize and leak individual training examples.

• Epsilon (ε): The privacy budget; lower values mean stronger privacy.
• DP-SGD: Clips per-sample gradients and adds Gaussian noise during local training.
• Prevents membership inference attacks on personalized models. This is essential for deploying personalized FL in regulated industries like healthcare and finance.