Fine-Grained Poisoning Attack to Local Differential Privacy Protocols for Mean and Variance Estimation

Explore a cutting-edge research presentation on data poisoning attacks against local differential privacy (LDP) protocols. Delve into the proposed fine-grained attack that manipulates mean and variance estimations in LDP systems. Learn about the novel output poisoning attack (OPA) technique, which injects fake data into the output domain of local LDP instances. Examine the security-privacy consistency observed in LDP and gain insights into the evolving threat landscape of data poisoning attacks. Compare the effectiveness of OPA against baseline attacks using real-world datasets. Discover a new defense method for recovering result accuracy from polluted data collections and understand implications for secure LDP design. This 20-minute conference talk from USENIX Security '23 offers valuable knowledge for researchers and practitioners in the fields of privacy-preserving data analysis and cybersecurity.