Style-aware adversarial pairwise ranking for image recommendation systems

The vulnerability of Machine Learning (ML) models to adversarial attack and their prominence pose security issues, notably in image recommendation systems. The adversarial training method is an excellent strategy for improving the generalization capacity of ML models by creating attacks in the embedding space during training. While there has been a plethora of testing on image recommendation system vulnerabilities and defenses, iterative adversarial training methodologies have received little attention. Furthermore, when browsing visual images, consumers are more interested in the content and how well the image style matches the content. However, when compared to image content, the impact of image styles on the adversarial recommendation community has rarely been examined. In this work, we propose a robust Adversarial Content and Style Bayesian Personalized Ranking (ACSBPR) approach that leverages content and style features for image recommendation. The ACSBPR technique makes three significant contributions: (1) Incorporate content and style features jointly for image recommendation. (2) Present a multi-objective pairwise ranking with Dynamic Negative Sampling to optimize the system and anticipate consumer preferences. (3) To reduce the influence of the attack, we train the ACSBPR objective function using minimax iterative adversarial training. Extensive investigations on the Flickr dataset demonstrate that our strategy achieves better performance when compared to state-of-the-art image recommendation models.