Secure Aggregation

Secure aggregation incurs significant computational and communication overhead, and its robustness is limited in cross-device scenarios facing frequent dropouts.

Secure aggregation can be stacked with differential privacy, the former hiding individual updates and the latter limiting privacy leakage of the aggregate result.

Secure Aggregation is a cryptographic protocol that allows multiple parties to keep their local model updates undisclosed while the server obtains only the aggregate sum of all updates.

The scheme does not require clients to use a uniform quantization level; clients can independently choose their quantization level according to their own bandwidth conditions without affecting the security and correctness of aggregation.

In federated learning, secure aggregation allows the server to compute the sum of the local model updates of N clients without leaking any individual client's update information (beyond the sum).

Existing heterogeneous-quantization secure aggregation methods leak partial aggregations of model updates, and under certain dropout and client-server collusion patterns may leak fragments of individual model updates.

PCHMR uses federated learning and a secure aggregation variant to train state-of-the-art HMR models, and evaluates their performance across a range of client scales and degrees of data heterogeneity.

Secure aggregation guarantees that the server cannot see the plaintext updates of any individual client and can only obtain the sum of all updates, thereby preventing user information from being inferred from updates.