Insider Collusion Attack on Distributed Machine Learning System and its Solutions - A Case of SVM

Among several known data breaching problems, insider collusion attacks have been one of the fastest-growing threats. However, there are very few insider-related researches involving the distribution of kernel-based data mining, such as distributed Support Vector Machine (SVM). Prior works often fall short by addressing a multifactorial model that is more limited in scope and implementation than addressing "insiders within an organization" colluding with outsiders. In this talk, I will introduce a novel insider collusion attack, in which a faulty system allows collusion to go unnoticed when an insider shares data with an outsider, who can then recover the original data from message transmissions (intermediary kernel values) among organizations. This attack requires only accessibility to a few data entries within the organizations rather than requiring the encrypted administrative privileges typically found in the distribution of data mining scenarios. I will also provide the analysis about the minimum amount of insider data necessary to launch the insider attack. Finally, I will discuss several proposed privacy-preserving schemes, like Differential Privacy, Homomorphic Proxy Re-Encryption, and etc., to counter the described attack.