Proactive Privacy-preserving Learning for Retrieval

Deep Neural Networks (DNNs) have recently achieved remarkable performance in image retrieval, yet posing great threats to data privacy. On the one hand, may misuse a deployed DNNs based system look up without consent. other organizations or individuals would legally illegally collect train high-performance models outside scope of legitimate purposes. Unfortunately, less effort has been made safeguard privacy against malicious uses DNNs. In this paper, we propose data-centric Proactive Privacy-preserving Learning (PPL) algorithm for hashing which achieves protection purpose by employing generator transfer original into adversarial with quasi-imperceptible perturbations before releasing them. When source is infiltrated, can confuse menacing retrieval make erroneous predictions. Given that prior knowledge not available, surrogate model instead introduced acting as fooling target. The framework trained two-player game conducted between and model. More specifically, updated enlarge gap data, aiming lower search accuracy contrary, opposing objective maintain performance. As result, an effective robust encouraged. Furthermore, facilitate optimization, Gradient Reversal Layer (GRL) module inserted connect two models, enabling one-step learning. Extensive experiments on three widely-used realistic datasets prove effectiveness proposed method.