Federated Data Valuation

The foundational principle of data valuation, measuring a client's worth by the performance delta observed when their dataset is added to a coalition of other clients. This isolates the unique predictive power of their local RF samples.

• Leave-One-Out (LOO): A naive baseline that measures the error increase when a single client is removed.
• Coalitional Game: Frames the valuation as a cooperative game where the 'payout' is model accuracy, and players are data owners.
• Orthogonality: High-value clients often possess data that is statistically orthogonal to the rest of the federation, filling critical coverage gaps.

A rigorous game-theoretic solution concept that fairly distributes the global model's performance gain among all participating clients. It computes the weighted average marginal contribution of a client across every possible permutation of coalition formation.

• Exhaustive Permutations: Requires evaluating model performance for all 2^N subsets of clients, which is computationally intractable for large federations.
• Monte Carlo Approximation: Practical implementations use random sampling of permutations to estimate Shapley values with bounded error.
• Fairness Guarantee: Satisfies axioms of efficiency, symmetry, dummy player, and additivity, ensuring no client subsidizes another.

An efficient approximation algorithm designed specifically for iterative gradient-based training, avoiding the prohibitive cost of full model retraining. It tracks the inner product of gradients to estimate value.

• Gradient Similarity: Values a client based on how much their local gradient updates align with the direction of the global model's improvement.
• One-Pass Computation: Estimates all client values during a single federated training run by recording intermediate gradient contributions.
• Complexity Reduction: Reduces computational overhead from exponential O(2^N) to linear O(N) in the number of training iterations.

A robust method that estimates a client's value by measuring the performance degradation when their data is replicated and removed from the federation, simulating a counterfactual world without their unique signal environment.

• Data Replication: Creates a synthetic baseline by duplicating a client's dataset across the federation before removal.
• Signal Diversity: Highly effective for RF applications where a client's value stems from capturing a rare interference pattern or channel condition.
• Robustness Check: Less sensitive to stochastic training noise than single-seed marginal contribution estimates.

The operational output of data valuation, translating Shapley values into tangible rewards to incentivize high-quality data contribution and mitigate free-riding in open federations.

• Proportional Rewards: Distributes monetary compensation, compute credits, or priority model access based on normalized valuation scores.
• Quality over Quantity: Rewards clients with rare, high-diversity RF datasets more than those contributing redundant samples.
• Sybil Resistance: Prevents a malicious actor from gaining undue influence by splitting a single dataset across multiple fake client identities.

A lightweight heuristic that scores clients based on their local model's accuracy on a global validation set held by the server, serving as a fast proxy for true Shapley value when computation is constrained.

• Loss Ranking: Clients whose local updates minimize the loss on a curated, balanced validation set are assigned higher value.
• Distribution Shift Detection: Low proxy scores can flag clients whose local RF data distribution has drifted significantly from the global task.
• Computational Efficiency: Requires no additional retraining or coalition sampling, making it suitable for resource-constrained edge orchestrators.

A game-theoretic concept adapted to calculate the marginal contribution of each client's local dataset to the global model's performance. It computes a weighted average of a client's performance impact across all possible coalitions of participants.

• Fairness Guarantee: Satisfies axioms of efficiency, symmetry, and additivity
• Computational Challenge: Exact calculation requires 2^N model retrainings, making it intractable for large N
• Approximation Methods: Truncated Monte Carlo sampling and gradient-based estimation (e.g., TMC-Shapley) reduce complexity
• Wireless Application: Identifies high-value spectrum sensors in a federated sensing network, enabling prioritized resource allocation

A lightweight alternative to Shapley value that estimates a client's data quality contribution by measuring the change in validation loss when their update is included versus excluded from aggregation.

• Leave-One-Out (LOO): Simplest form, measuring performance delta when a single client is removed
• Influence Functions: Trace model prediction changes back to individual training samples without retraining
• Gradient Similarity: Compare a client's update direction with the aggregate update to assess alignment
• Practical Use: Enables real-time client selection in cross-device FL where Shapley computation is infeasible

Pre-training assessment of a client's local dataset to predict its potential value before federated training begins, reducing wasted computation on low-quality participants.

• Label Accuracy: Detect noisy or mislabeled samples through cross-validation with a trusted validation set
• Distribution Coverage: Measure how well a client's data covers the global feature space using maximum mean discrepancy (MMD)
• Information Content: Quantify the entropy or mutual information of a client's dataset relative to the learning task
• RF-Specific Metrics: Signal-to-noise ratio (SNR) distribution and modulation diversity in spectrum datasets

The economic framework that translates data valuation scores into tangible rewards, motivating high-quality clients to participate honestly in federated training.

• Contribution-Based Rewards: Allocate monetary payments or computational credits proportional to Shapley values
• Reputation Systems: Maintain a non-tamperable ledger of historical contribution scores to build long-term trust
• Reverse Auction: Clients bid their required compensation, and the server selects the optimal subset under a budget constraint
• Game-Theoretic Stability: Ensure mechanisms are strategy-proof so that truthful reporting of data quality is the dominant strategy

Techniques to identify clients that submit low-quality or random model updates to receive the global model without contributing meaningful data, a direct application of data valuation.

• Update Norm Analysis: Detect abnormally small or random gradient norms that indicate no real training occurred
• Cosine Similarity Thresholding: Flag clients whose update direction is consistently orthogonal to the aggregate
• Validation Baiting: Send a held-out validation task to verify that a client's update improves performance on known data
• RF Context: Detect malicious sensors injecting noise into a federated automatic modulation classification system

A statistical method from robust statistics that estimates data value by measuring how model performance degrades when a client's data is replaced with a random replica sampled from the population distribution.

• Infinitesimal Jackknife: Approximates the effect of removing a client using first-order gradient information
• Data Shapley via Gradient Descent: Reformulates Shapley value computation as an optimization problem solvable with SGD
• Scalability Advantage: Reduces computation from exponential to polynomial time for certain model classes
• Wireless Edge Use: Efficiently values intermittent IoT sensor contributions in cross-device federated learning