Recommendation systems are widely used in domains such as e-commerce, healthcare, and entertainment to alleviate information overload. However, they remain vulnerable to poisoning attacks, where adversaries inject crafted fake profiles to manipulate recommendation outcomes. This paper investigates such attacks—also known as profile or configuration attacks—and proposes BEGAN (Bilateral Generator Evaluation Generative Adversarial Network) to enhance attack effectiveness and realism. BEGAN includes four modules: sampling, bilateral generator, evaluation, and discriminator. Unlike conventional models that treat unlabeled data as negative, BEGAN introduces both positive and unlabeled sampling strategies. It further improves upon traditional GANs by dividing the generator into a user generator and an item generator, aligning with the user-item rating nature of recommendation data. An evaluation module is incorporated to assess how generated fake profiles impact the system, forming a triple adversarial training structure with the generator and discriminator. Experiments on two real-world datasets across five evaluation metrics demonstrate that BEGAN not only outperforms ten baseline models but also generates fake profiles that are significantly harder to detect, highlighting its effectiveness in modeling sophisticated poisoning attacks.