Data poisoning attacks against local differential privacy triangle counting protocols
Data Poisoning Attacks to Locally Differentially Private Triangle Counting Protocols
AI summary
Local differential privacy (LDP) triangle-counting protocols are vulnerable to data poisoning attacks: small, protocol-compliant local manipulations by malicious users can be systematically amplified into large global estimation bias through reconstruction and calibration. The paper analyzes the attack surface of both non-interactive and interactive protocols, proposes a unified gain-driven framework, and instantiates it on representative protocols. Experiments on four real graph datasets show the attacks significantly distort triangle-count estimates, and existing and simple defenses are insufficient to mitigate them.