Yuval Elovici

Feature masking, the practice of withholding sensitive attributes from potential adversaries, is widely relied upon as a frontline defense against privacy attacks on tabular machine learning (ML) models. However, this paper demonstrates that the protection it offers is an illusion. While traditional Membership Inference Attacks (MIAs) assume adversaries possess complete knowledge of a target's features, we formalize a highly realistic threat model where attackers observe only partial records. To exploit this, we introduce the Missing Features-Membership Inference Attack (MF-MIA), a practical two-stage black-box framework. MF-MIA first imputes with-held attributes using auxiliary data with overlapping features, and subsequently mounts a shadow-model-based MIA on the completed queries. In a cross-database case study (target model trained on the General Social Survey; auxiliary data from UCI Adult), MF-MIA achieves up to 59.3% membership accuracy even when a sensitive feature is entirely withheld. By benchmarking multiple completion strategies, including statistical priors, model-based imputation, TVAE, and masked autoencoders, we show that attack power scales with imputation quality, independent of target-model feature importance. Finally, we demonstrate that Large Language Models (LLMs) can effectively substitute for tabular auxiliary data in data-scarce environments, decisively proving that feature masking cannot stand alone as a robust privacy defense.