researchsquare.comSource publication:
Loading content…
Sports images sit with schools, clubs and other organizations, and federated learning avoids sending raw pictures but still transmits model updates that can leak information, while the clipping and noise of differential privacy damage useful knowledge. FedSS scores each network stage for how shareable it is before clipping, uploads only the perturbed cross-client sport semantics, and keeps client-specific appearance and background context local. On sport22 and sport100 it beats the strongest baseline DPSIR-Fed by 5.78 and 5.71 points under visual-context heterogeneity, and raises sport22 accuracy from 42.67% to 46.81% at the strong-privacy setting ε=4.
Federated sports recognition can now share only the part of a model update that generalizes across clients, instead of clipping and perturbing the whole update uniformly. Earlier differentially private federated learning treated local representations and model updates as one whole, without asking which components suit cross-client sharing. Under visual-context heterogeneity on sport22 and sport100, FedSS reaches Top-1 accuracy 5.78 and 5.71 percentage points above the strongest baseline DPSIR-Fed; when label and visual heterogeneity combine, the margins are 3.11 and 2.64 points.
The ResNet-18 backbone is split into five stages, each given a shareability coefficient between 0 and 1 that softly separates a candidate update into a sport-semantic part and a client-context part. Each coefficient averages three equally weighted criteria: the stage's contribution to correct sport recognition, its consistency with the server's cross-client update history, and the stability of its signal under privacy perturbation. Only the sport-semantic part is L2-clipped and Gaussian-perturbed before upload, while the context part stays as a local residual that is never communicated; experiments cover 100 clients with 10% participation and 500 clients with 2% participation, averaged over three random seeds.
The demonstrated scope is a ResNet-18 classifier on the two public sports image benchmarks sport22 and sport100, covering IID, label heterogeneity, visual-context heterogeneity and their combination, at 100 clients with 10% participation and 500 clients with 2% participation. Defaults are 400 communication rounds, 20 local epochs, clipping threshold 0.4 and noise multiplier 0.95, corresponding to roughly ε=12 under S1. The authors propose extending to video and multimodal sports data, naturally occurring cross-institutional heterogeneity, and finer-grained adaptive shareability estimation.
The privacy guarantee relies on every released client update being independently clipped and perturbed, with shareability coefficients and local residuals never uploaded; the semantic-context decomposition in the analysis is used only for analysis and is not available during training. The shareability threshold τ and context-retention coefficient μ work best at moderate values, and both too small and too large values weaken the balance; absolute accuracy of all methods drops when the federation grows from S1 to S2. The authors state that code and experiment configurations will be released on reasonable request or after publication.