Boosting deep cross-modal retrieval hashing with adversarially robust training

Deep hashing methods effectively enhance the performance of conventional machine learning retrieval models, particularly in visual medium evolving cross-modal retrieval tasks, by relying on the outstanding feature extraction ability of deep neural networks (DNNs). The state-of-the-art deep hashing research focuses on designing prominent models by employing DNNs to discover semantic information from different modalities of data and execute relevant information retrieval tasks. However, the robustness attribute considered essential for reliable DNN model design has limited concerns on deep hashing models. In this article, we present an end-to-end adversarial training framework for cross-modal retrieval. Our framework leverages a projected gradient descent(PGD)-based method to generate adversarial samples, which are then combined with normal samples to achieve robust training. Our approach addresses the vulnerability issues of existing cross-modal retrieval models and fills the gap in retrieval task design. We conduct extensive experiments and compare our model with state-of-the-art cross-modal retrieval models on three benchmark datasets to verify that our model can effectively boost the performance of deep hashing retrieval models on cross-modal retrieval . This work highlights the effectiveness of adversarial training in efficient deep hashing model design.