GAN-Based Domain Inference Attack

Model-based attacks can infer training data information from deep neural network models. These heavily depend on the attacker's knowledge of application domain, e.g., using it to determine auxiliary for model-inversion attacks. However, attackers may not know what model is used in practice. We propose a generative adversarial (GAN) based method explore likely or similar domains target -- domain inference (MDI) attack. For given (classification) model, we assume that attacker knows nothing but input and output formats use derive prediction any desired form. Our basic idea affect GAN process candidate domain's dataset easy obtain. find distort procedure less if more domain. then measure distortion level with distance between GAN-generated datasets, which be rank model. experiments show an MDI top-ranked effectively boost result