Adaptive information-constrained mapping for feature compression in edge AI and federated systems

This article explores the problem of efficient feature compression in distributed intelligent systems with limited resources, particularly within the context of Edge AI and Federated Learning. The relevance of this study is driven by the growing need to reduce communication overhead under conditions of unstable Quality of Service, limited bandwidth, and high heterogeneity of input data. The scientific novelty lies in the development of a consistent entropy-regularised compression model that combines variational latent mapping, non-negativity-constrained projection design, and stochastic-Boolean transformation of the feature space. A generalised compression quality functional is proposed, integrating the directed Kullback-Leibler divergence, an entropic regularisation component, and a guarantee of preserving the semantic relevance of the compressed representation. Efficient projection-gradient optimisation algorithms have been developed, suitable for implementation in constrained computational environments. The practical effectiveness of the approach has been confirmed through experiments on the HAR and PAMAP2 datasets: a 6-eightfold reduction in entropy load was achieved while maintaining classification accuracy above 94% and a high level of semantic fidelity in the reconstructed data. The models were deployed on low-power devices (Jetson Nano, Raspberry Pi 4), where they demonstrated robustness to noise and loss, as well as superiority over current SOTA solutions (FedEntropy, EDS-FL, SER) in terms of compression efficiency, adaptability to heterogeneous distributions, and stability under unstable transmission conditions.