RAIN: Secure and Robust Aggregation under the Differentially Private Shuffle Model

The RAIN framework achieves secure aggregation that balances privacy, robustness and verifiability under the differentially private shuffle model (Shuffle-DP). It uses symbolic-space aggregation to limit the impact of malicious poisoning, designs a shuffle and aggregation protocol based on additive secret sharing, maintains tight privacy guarantees, and verifies the aggregation result each round to detect selective dropping. Experiments show it is robust to poisoning attacks, with almost no loss in accuracy and convergence, communication cost reduced by up to 90 times compared with prior work, and aggregation speed improved by 10 times.