Secure Aggregation

A cryptographic scheme enabling computation directly on ciphertexts. In secure aggregation, additively homomorphic encryption allows the server to sum encrypted model updates without ever decrypting individual contributions. The result, when decrypted, matches the sum of the plaintexts. Partially homomorphic encryption (PHE) supports only one operation type and is computationally lighter than fully homomorphic encryption (FHE), making it practical for federated averaging workflows.

A mathematical framework that provides provable privacy guarantees by injecting calibrated noise into data or model updates. When combined with secure aggregation, differential privacy bounds the information leakage from the aggregated result itself. Key parameters:

• Epsilon (ε): Privacy budget controlling the trade-off between privacy and accuracy
• Delta (δ): Probability of catastrophic privacy failure
• Local DP: Noise added by each client before encryption
• Central DP: Noise added by the server post-aggregation

The property of a distributed system to function correctly despite arbitrary or malicious failures among participants. In secure aggregation, Byzantine fault tolerance ensures the protocol remains robust when:

• Malicious clients submit crafted updates to corrupt the global model
• Clients drop out mid-protocol due to network failures
• Adversarial nodes attempt to prevent honest clients from contributing Techniques like Krum, trimmed mean, and multi-Krum aggregation rules filter out anomalous updates before summation.

A cryptographic method for distributing a secret among a group of participants, where only qualified subsets can reconstruct it. In secure aggregation protocols like Google's SecAgg, each client's model update is split into secret shares distributed to other clients. The server can only reconstruct the sum after aggregating shares, never an individual update. Shamir's Secret Sharing and additive secret sharing are common schemes. This approach tolerates client dropouts without compromising privacy.

A hardware-enforced secure enclave within a processor that guarantees confidentiality and integrity of code and data. In secure aggregation, a TEE acts as a trusted intermediary:

• Decryption and aggregation occur entirely within the enclave
• The host OS, cloud provider, and other tenants cannot inspect the data
• Remote attestation cryptographically verifies the enclave's integrity to clients Intel SGX, AMD SEV, and ARM TrustZone are production TEE implementations. This approach shifts trust from cryptography to hardware.

A privacy attack where an adversary reconstructs representations of private training data from model parameters or gradients. Secure aggregation directly mitigates this threat by preventing the server from inspecting individual gradient updates. Without secure aggregation, an honest-but-curious server could:

• Reconstruct training images from gradient updates in vision models
• Extract memorized text sequences from language model gradients
• Infer membership of specific data points in the training set Gradient leakage is a primary threat vector that secure aggregation neutralizes.