With their poster, the Fraunhofer IDMT scientists presented their FLCrypt approach for the first time, a tool for improving data protection for Federated Learning tasks using Fully Homomorphic Encryption.

The aim of the IEEE 5th International Symposium on the Internet of Sounds (IEEE IS² 2024) is to bring together academia and industry to explore and advance the development of Internet of Sounds technologies using novel tools and processes.

The method ensures that the central server cannot see the training participants' unencrypted model data. The approach was tested with a special audio task in which metal balls with different surfaces, one of which is damaged, were to be identified by their sounds – a scenario representative for the detection of flaws in industrial sound analysis.

The encryption achieved comparable accuracy to the unencrypted scenario but led to a substantial increase in runtime and memory use. Despite these challenges, the results show that FLCrypt is a promising option for secure Federated Learning for sound classification tasks.

We congratulate our colleagues on this success!