Bounding information leakage in machine learning

Recently, it has been shown that Machine Learning models can leak sensitive information about their training data. This leakage is exposed through membership and attribute inference attacks. Although many attack strategies have proposed, little effort made to formalize these problems. We present a novel formalism, generalizing setups previously studied in the literature connecting them memorization generalization. First, we derive universal bound on success rate of attacks connect generalization gap target model. Second, study question how much stored by algorithm its set bounds mutual between attributes model parameters. Experimentally, illustrate potential our approach applying both synthetic data classification tasks natural images. Finally, apply formalism different strategies, with which an adversary able recover identity writers PenDigits dataset.