Federated learning (FL) has recently been adopted to train shared models across industrial Internet of Things (IoT) devices without revealing their private raw data. Conventional FL usually relies on a central server for coordination. However, in reality, the central server is not fully trusted, which means that it may be collecting data, raising concerns about data leakage and misuse. Here, we propose a decentralized FL algorithm based on deep neural networks to address the problem of untrusted central servers. The original problem is decomposed into several subproblems with consensus constraints, which can be solved by local computation and communication. The proposed algorithm combines the decentralized average consensus and alternating direction method of multipliers. Several decentralized algorithms are employed for comparison, and the issue of heterogeneous data is discussed. Experimental evaluation shows the effectiveness of our proposed algorithm.

In the context of the current rapid development of the financial market, how to establish an effective stock price index model to avoid investment risks and enhance investment returns for investors has become a subject of great concern. On this basis, a new method of stock price index prediction employing deep neural networks is investigated. This paper employs the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) technique to break down the time series of the stock index into its constituent Intrinsic Modal Functions (IMFs). According to the similarity of the values of Fuzzy Entropy (FE), the subsequence is reorganized to become a new sequence, which highlights the fluctuation state of the stock index at different frequencies and improves the forecasting efficiency. In terms of a forecasting method, the combination of Temporal Convolutional Network (TCN), Gated Recurrent Unit (GRU), and Convolutional Block Attention Module (CBAM) is used to forecast the reorganized subsequence, and the final forecast results are reconstructed to obtain the final prediction value. To further evaluate the performance of the proposed CEEMDAN-TCN-GRU-CBAM model, this paper selects four representative stock indices in emerging and developed markets while comparing them with the benchmark model. It uses four evaluation metrics to measure the model performance. The study shows that the proposed model outperforms other benchmark models has better robustness and universality, and has higher forecasting accuracy.

Combined optimization of grid-search LSSVR modelling with SNV-SGS pre-processing improves the predictive abilities of the FT-MIR analysis of human blood HGB.