Defenses against poisoning attacks under Shuffle-DP
Defense against Poisoning Attacks under Shuffle-DP
AI summary
The researchers propose the first general defense framework against poisoning attacks on Shuffle-DP for all retained union queries, which can convert any Shuffle-DP protocol into a poisoning-resistant version. The framework maintains asymptotically equivalent error when there is no attack, and the error increases only by a polylogarithmic factor when a constant number of attackers exist. Experiments on sum, frequency estimation and range counting queries show that it effectively defends against poisoning while maintaining utility and communication efficiency.