Federated learning (FL) has gained significant attention as a distributed machine learning approach in recent years. However, in the context of the Internet of Things, terminal devices exhibit considerable heterogeneity in data distribution, communication, and computational resources. This heterogeneity can introduce bias, hindering model accuracy through iterative training. Moreover, disparities in communication and computational capabilities limit FL to resource-constrained devices, slowing convergence. To address these challenges, we propose FedDG, an FL framework that integrates model compression and dynamic fusion. Specifically, for nonindependently and identically distributed data, we adaptively adjust the client compression rate based on model training errors, reducing transmission costs while improving both convergence speed and accuracy. Furthermore, clients are dynamically selected for aggregation in each round based on training time and model performance, alleviating the impact of low-resource clients and reducing communication overhead. Experimental results show that FedDG significantly lowers communication bandwidth and data transmission size, and accelerates model convergence under data heterogeneity compared to benchmark schemes, such as FedAvg, FedNova, STC, DF_FL, DAdaQuant, and adaptive quantization in device selection strategy (AQUILA).

Cross-domain Federated Learning (FL), as an emerging paradigm of distributed machine learning, enables collaborative training across multiple parties while preserving data privacy. However, cross-domain data alignment, as a critical step in FL, often faces the challenge of private data leakage. Applying the Threshold Private Set Intersection (TPSI) protocol to data alignment in cross-domain FL proves to be a secure and efficient solution. Nevertheless, existing two-party or multi-party TPSI protocols commonly rely on high-overhead public-key cryptographic algorithms such as Homomorphic Encryption (HE), resulting in low operational efficiency. In response to these limitations, this work introduces a computationally efficient two-party and multi-party TPSI protocols based on Oblivious Key-Value Store (OKVS) and Threshold Secret Sharing (TSS). To determine whether the intersection cardinality meets the threshold, OKVS is employed to associate elements with secret sub-shares, effectively preventing leakage of intersection cardinality. Additionally, to avoid the exponential combination of secret shares, this paper adopts a Reed-Solomon decoding algorithm based on Fast Fourier Transform (FFT) for secret reconstruction. Furthermore, building upon the two-party TPSI protocol, we design an efficient multi-party TPSI protocol by incorporating unconditional zero-sharing. Finally, a detailed correctness proof and performance analysis of the two protocols demonstrate that the proposed solution is secure, feasible, and applicable to cross-domain FL data alignment scenarios involving varying numbers of participants.

Conventional two-dimensional structured illumination microscopy (SIM) requires 9 raw images to reconstruct a super-resolved image. In order to increase the frame rate of 2DSIM, attempts are being made to reduce the number of raw SIM images. However, all the proposed SIM reconstruction algorithms (SIM-RA) capable of reconstructing super-resolution (SR) image with a reduced number of raw SIM images operate in the spatial domain. Here, we present a frequency domain SIM-RA based on ordinary least squares technique, which enables reconstruction of SR image using 4 raw SIM images. Unlike the spatial domain RA, which produces the SR image through iterative convergence, the presented RA provides a single step solution. It also reveals the fundamental limitation of least number of raw images required for resolution doubling in SIM.