Secure aggregation stays information-theoretically private with heterogeneous client quantization
The claim
The authors assert their scheme achieves information-theoretic security against up to T colluding semi-honest clients and the server while supporting heterogeneous quantization.
Published as: “Information-Theoretic Secure Aggregation With Heterogeneous Client Quantization”
So what
Enables privacy-preserving federated learning across devices with differing bandwidths, but practical deployment may be limited by accuracy instability and the need for reliable secure links.
The five-minute version
The paper presents an information-theoretic secure aggregation protocol that tolerates up to T colluding clients and D drop-outs while allowing each client to use its own quantization level. Experiments on MNIST with 10 clients show test accuracy comparable to FedAvg, but heterogeneous quantization introduces larger scaling factors and noticeable accuracy fluctuations. The evidence is limited to a small public benchmark and a single model architecture.
What they found 1
- The protocol achieves test accuracy similar to FedAvg despite quantization.
“Our protocol provides similar accuracy as that of FedAvg, while quantization errors affect the training stability.”
What undercuts it the paper’s own results
- Accuracy fluctuates more in heterogeneous settings due to the scaling factor L, which introduces error in the estimated aggregate.
Stated limits 4
- We implement our secure aggregation protocol on the MNIST dataset [19] for N = 10 clients, in an i.i.d. setting.
- The fluctuations in accuracy are due to the scaling factor L which introduces error in θΣ,est.
- We assume that the clients can establish error-free, secure communication links with each other and with the server.
- the server and clients are semi-honest, i.e., they execute the protocol truthfully, but attempt to infer θn from the messages they receive.
What’s new here 4
- Supports heterogeneous client quantization levels within a single secure aggregation round.
- Provides information-theoretic privacy against up to T colluding clients and a semi-honest server.
- Handles up to D client drop-outs without compromising correctness.
- Uses a black-box (T+1, N) secret sharing scheme on the torus with integer-valued coefficients independent of the underlying group.
Threat model stated by the source
Artifacts & methods 5
- black-box (T+1, N) threshold secret sharing on the torussystem
- MNIST datasetdataset
- FedAvg
- stochastic rounding quantization
- normalization scaling factor L