Robust Defense Mechanisms for Agentic News Recommenders: Mitigating Data Poisoning Attacks

As personalized content delivery becomes central to modern news consumption, agentic news recommenders face rising threats from data poisoning attacks. These attacks inject malicious data into the training pipelines, distorting recommendations, spreading misinformation, and eroding user trust.In this study, we introduce an integrated defense framework that combines robust learning algorithms, data sanitization techniques, and anomaly detection methods to safeguard news recommendation systems. Through extensive experiments on a widely used real-world news dataset, our framework demonstrated significant improvement in both accuracy and resilience, achieving a 10.6% increase in NDCG@10 under attack conditions.We also explore the ethical dimensions of deploying such defenses, emphasizing the balance between fairness, transparency, and robustness. Our findings contribute to the development of secure, trustworthy, and adaptable news recommender systems capable of withstanding evolving adversarial threats.