“Why is the Sky Blue?” – On the Feasibility of Privacy-Friendly Conversational LLM Smart Toys
摘要
The integration of large language models into smart toys introduces significant privacy risks for children due to the transmission of data to cloud servers for processing. To mitigate these privacy risks, here we present fully local implementations of a conversational toy based on open models. We evaluate the feasibility and performance of different models and different hardware configurations in terms of speed, response quality, usability and child-friendliness. Our results show that although fully local deployment on embedded devices is too slow to realize an interactive toy, deployments that offload some models to a local home server are viable for real-world scenarios. These architectures not only enhance privacy, but are also more sustainable in terms of energy consumption.