Model Inversion Attack: Analysis under Gray-box Scenario on Deep Learning based Face Recognition System

In a wide range of ML applications, the training data contains privacy-sensitive information that should be kept secure. Training systems by makes model inherent to data. As structure has been fine-tuned data, can abused for accessing estimation in reverse process called inversion attack (MIA). Although, MIA applied shallow neural network models recognizers literature and its threat privacy violation approved, case deep learning (DL) model, efficiency was under question. It due complexity DL structure, big number parameters, huge size registered users thereof class labels. This research work first analyses possibility on recognition system, namely face recognizer. Second, despite conventional white box scenario having partial access users' non-sensitive addition is implemented system just parameters but not any user information. this aspect, it semi-white or other words gray-box scenario. The experimental results targeting five CNN-based approve regeneration images even gray some evaluation score low generated are easily recognizable, high facial features targeted identities observable. objective subjective evaluations demonstrate cyber-attack only feasible also serious with increasing alert state future as there considerable potential integration more advanced techniques MIA.