Uncovering Memorization in Timeseries Imputation models: LBRM Membership Inference and its link to attribute Leakage

Researchers have identified critical privacy vulnerabilities in deep learning models used for time series imputation, demonstrating that these models can leak sensitive training data through membership and attribute inference attacks. The study introduces a two-stage attack framework that successfully retrieves significant portions of training data even from models designed to be robust against overfitting-based attacks.

• →Time series imputation models in healthcare, IoT, and finance are vulnerable to black-box inference attacks that can expose sensitive training data.
• →A novel two-stage attack framework combines membership inference attacks with the first attribute inference attack specifically designed for timeseries models.
• →The proposed membership attack achieves significantly higher detection accuracy than baseline methods, with a tpr@top25% score substantially above naive attack baselines.
• →Models trained from scratch and fine-tuned models with accessible initial weights are both susceptible to these privacy attacks.
• →The membership attack provides strong predictive capability for attribute inference success, achieving 90% precision compared to 78% in general cases.