SoK: Colluding Adversaries in Machine Learning Pipelines

Researchers have identified security risks in machine learning models due to potential collusion among adversaries, and proposed a framework to study these threats. A new programming model, SemPipes, aims to improve ML pipeline development.

Machine learning models are vulnerable to security, privacy, and fairness risks, according to a study published on arXiv^[1]. The research found that adversaries with different characteristics can collude to amplify attacks. Existing work lacks a systematic framework to explore this collusion, the study noted. To address this, the researchers proposed a framework to study the implications of adversaries' characteristics on collusion. Meanwhile, another study on arXiv^[2] highlighted the challenges of developing ML pipelines, which require extensive data preparation and feature engineering. The researchers introduced SemPipes, a novel programming model that extends ML pipelines with declarative, LLM-powered semantic data operators. SemPipes synthesizes specialized implementations of semantic operators at pipeline training time, conditioned on dataset characteristics and pipeline context^[2].