292 / 2024-04-24 10:39:54

Assessing and Enhancing Adversarial Robustness for Review-Based Recommender Systems: A Design Science Approach

adversarial robustness,recommender systems,online reviews,design science

Recommender systems (RS) are widely adopted in e-commerce platforms, yet they have faced serious adversarial attacks in recent years. While existing studies have focused on assessing the robustness of RS when the interaction matrix and image are manipulated, the vulnerability of review-based RS (R-RS) when online reviews are manipulated remains unexplored. In this research, we adhere to the principles of adversarial robustness theory and propose a novel assessment and enhancement framework for R-RS. We adopt the computational design science paradigm to develop a novel assessment method, which includes reconnaissance and execution phases, and reveals the vulnerability of R-RS by creating adversarial samples. We then design improvement methods for each phase. In the reconnaissance phase, we introduce a stochastic recommendation process that increases the difficulty of gathering model information. In the execution phase, we propose weighted input dropout and weighted adversarial contrastive learning to ensure that R-RS perceives undistorted information from adversarial samples, both in a certified and empirical manner. We thoroughly evaluate the performance of the proposed framework using ground truth data, and the results show that R-RS is indeed vulnerable to adversarial attacks, and the proposed methods significantly enhance its robustness.