Security issues in federated learning systems

The paper discusses key security problems in federated learning systems: protecting the privacy of participants' data from gradient inversion attacks and ensuring model resistance in the presence of poisoning attacks. A review of current approaches to defense against the above threats is presented, and limitations in attempting to apply them together are identified. Based on the analysis, we formulate our own ideas for further research aimed at developing more effective and balanced defense methods that consider both data privacy and poisoning attack resistance.