Large language models deployed in RetrievalAugmented Generation systems face intensifying regulatory mandates to support the “right to be forgotten” under frameworks such as GDPR and HIPAA, while simultaneously confronting emerging ML security threats including adversarial retrieval poisoning, prompt injection, and membership inference attacks. Conventional parametric unlearning approaches that require full or partial model retraining impose prohibitive computational costs for production-scale LLMs, creating an urgent need for efficient, secure deletion mechanisms. This paper introduces RAG-Forget, a unified behavioral unlearning framework that achieves effective data deletion through external memory modification rather than parameter updates, fundamentally decoupling privacy compliance from model retraining while hardening the system against adversarial exploitation. We formalize the RAG unlearning problem by defining three complementary evaluation metrics: Unlearning Success Rate (USR), Utility Retention (UR), and Robustness Score (RS), the latter specifically designed to measure resilience against ML security threats including paraphrasing attacks, indirect reasoning exploits, adversarial document injection, and retrieval poisoning. Our framework synergistically combines unlearned knowledge construction, retrieval policy updates, and prompt-level constraints to achieve 98.1% unlearning success while maintaining 95.2% utility retention, all within a security-hardened architecture. Comprehensive robustness evaluation demonstrates that RAG-Forget achieves 89.3% robustness against adversarial attacks compared to 62.3% for corpus deletion baselines, validating its dual efficacy for privacy compliance and ML security. Experiments across medical, legal, and user privacy domains confirm practical applicability in high-stakes regulatory environments. We release our evaluation framework and benchmark datasets to advance reproducible research in privacy-preserving and security-aware RAG systems, establishing that effective unlearning and ML security are not competing objectives but mutually reinforcing requirements for trustworthy deployment.

Machine Unlearning,Retrieval-Augmented Generation (RAG),Large Language Models (LLMs),Privacy Preservation,Right to be Forgotten,GDPR Compliance,Adversarial Robustness,ML Security,Membership Inference Attacks,Retrieval Poisoning