Maximizing Privacy in Reinforcement Learning with Federated Approaches
Shiva Mehta, Sumeet Singh Sarpal · 2024
Data protection issues and security in the age of technological advancement have become problematic, mainly because these methods are increasingly widespread. This paper covers critical problems of Federated Reinforcement Learning (FRL) and shows how this technique could provide a viable solution to the issue. We are attempting the use case scenario of the Federation Q-learning in the Labyrinth difficulties here. It ensures that technological requirements are adapted at all levels and keeps the information secret. A trial has been made to test the application of IRL on AI training, enabling the scientists to acquire actual proof of the advantages of FRL against the theoretical approaches in training, which are the majority. The main quantitative facet of this case uses pretty scientific statistics. Research results from conventional Q-learning and our variation, codel bean Q-learning, are significant in drawing conclusions and gaining insights. The federated q-learning presented scores around 25% higher than the q-learning with an agent, which is one more advantage than the q-learning using an agent. Just as Federated Q-learning obtains fast policy optimization by reducing convergence time by 2 5%, parallelism also brings about quicker and more effective exploration of the Q-table. Federated Q-learning is no exception to the general rule that this model has an advantage in ensuring data privacy, yielding an evident four-point improvement in the data privacy score. This shows that this technology quickly becomes a trusted ally and can anonymize data without compromising users’ privacy and security. Moreover, in the final stage of our work, the implementation of the model pretty soon caused an improvement in computer efficiency by 40%. The results showed that the current traditional AIL techniques suddenly faced five times more risks concerning privacy and security.