RoLLMRec: a robust LLM-based recommender system for defending against shilling and prompt injection attacks

Large Language Models (LLMs) are increasingly being integrated into recommender systems, offering contextual reasoning, cross-domain adaptability, and natural language interaction. However, their adoption also introduces vulnerabilities such as prompt injection, semantic poisoning, and shilling attacks, which can distort recommendations and erode user trust. Addressing these risks is essential for the safe deployment of LLM-based recommenders. We propose RoLLMRec, a defense oriented architectural framework and evaluation methodology for LLM-based recommender systems that integrates prompt filtering, retrieval augmented grounding, trust aware scoring, and an auditing feedback loop. RoLLMRec improves robustness under the evaluated prompt level and semantic adversarial settings, while multimodal support is included at the architectural level only and is not empirically evaluated in the current experimental setup.RoLLMRec unifies five core components: (1) prompt shielding and input filtering to detect and block adversarial instructions; (2) retrieval-augmented generation to enrich factual grounding and reduce hallucination; (3) multimodal LLM encoding for text, metadata, and image inputs; (4) trust-aware scoring and Top-K ranking; and (5) adaptive feedback loops for continual learning. Evaluations on benchmark datasets such as Yelp, MovieLens, and Amazon Books show that RoLLMRec surpasses BERT4Rec, RecVAE, and LightGCN, improving NDCG@10 and HR@10 by up to 6% and 5%, respectively. Under a 10% prompt-injection attack, it maintains a Robust Hit Rate (RHR@10) above 0.63 and a Perturbation Sensitivity Index (PSI) below 0.135, achieving 15%–25% higher resilience. It also sustains a Semantic Stability Score (SSS) above 0.60 in zero-shot cross-domain transfer, confirming stable semantic intent.