Secure Aggregation with Homomorphic Encryption

MaskCrypt

MaskCrypt is a secure federated learning system based on homomorphic encryption. Instead of encrypting all the model updates, MaskCrypt encrypts only a subset of them to balance the tradeoff between security and efficiency. In this example, clients only select 5% of the model updates to encrypt during the learning process. The number of encrypted weights is determined by encrypt_ratio, which can be adjusted in the configuration file. A random mask will be adopted if random_mask is set to true.

cd examples/secure_aggregation/maskcrypt
uv run maskcrypt.py -c maskcrypt_MNIST_lenet5.toml