Editing Federated Learning (section)

== 5.4.3 Homomorphic Encryption and SMPC ==

Homomorphic Encryption (HE) supports computations on encrypted data without the need for decryption. In FL, a homomorphically encrypted gradient can be aggregated securely by the server, preventing it from seeing cleartext updates. This approach, however, introduces higher computational overhead, which can be burdensome for resource-limited IoT edge devices.<sup>[3]</sup>

Secure Multi-Party Computation (SMPC) is a related set of techniques that enables multiple parties to perform joint computations on secret inputs. In the context of FL, SMPC allows clients to compute sums of model updates without revealing individual updates. Although performance optimizations exist, SMPC remains challenging for large-scale models with millions of parameters.<sup>[1]</sup><sup>[5]</sup>