The new form of artificial intelligence training, i.e. federated learning, is becoming more popular in the last few years. It is an optimization problem that includes additional mechanisms such as aggregation and data transmission. In this paper, we propose a hybridization of this type of training with a meta-heuristic. The meta-heuristic algorithm is adapted to manage the entire process as well as to analyze the best models to minimize attacks on this type of collaboration. The proposed solution is based on minimizing the general model error, with additional control mechanisms for incoming models, or adapting the aggregation method depending on the quality of the model. The innovative solution has been analyzed in terms of its application to the problem of image classification using classical and convolutional neural networks, and the most popular meta-heuristic algorithms. The proposal was analyzed in terms of the accuracy of the general model as well as for security against poisoning attacks. We reached 91% of accuracy using the proposed method with the Red Fox Optimization Algorithm and 95% in terms of detection of poisoned samples in the database.

人工智能训练的新形式，即联邦学习，在最近几年变得越来越流行。它是一个优化问题，包括额外的机制，如聚合和数据传输。在本文中，我们建议将这种类型的训练与元启发式混合。元启发式算法适用于管理整个过程以及分析最佳模型以最大程度地减少对此类协作的攻击。建议的解决方案基于最小化一般模型错误，对传入模型具有额外的控制机制，或根据模型的质量调整聚合方法。已经分析了创新解决方案在使用经典和卷积神经网络的图像分类问题中的应用，以及最流行的元启发式算法。该提案从通用模型的准确性以及针对中毒攻击的安全性方面进行了分析。我们使用所提出的方法和 Red Fox 优化算法达到了 91% 的准确率，在检测数据库中的中毒样本方面达到了 95%。