<p>The proliferation of misinformation and rumors on social networking platforms poses significant challenges to online credibility, information integrity as well as public trust. Various conventional rumor detection approaches often struggle with linguistic ambiguity, data sparsity as well as subtle variation in rumor content. In order to address such shortcomings, this paper proposes a novel hybrid framework named deep bidirectional generative adversarial network (DBi-GANet) which is the integration of deep Bi-LSTM model and generative adversarial network. As a preliminary phase, a robust data preprocessing pipeline is utilized to enhance the consistency and quality of the social media text. This includes removal of hashtags, stop words, URLs followed by text normalization, tokenization, and word embedding using pretrained GloVe vectors. In the proposed DBi-GANet model, the generator is trained to produce realistic rumor text sequences that simulate both linguistic and structural patterns of social media rumors. The discriminator then modeled using Bi-LSTM that learns to classify sequences as rumor or non-rumor while capturing temporal dependencies from word contexts. The adversarial training enhances the robustness of the discriminator against deceptive and synthetic rumor content, allowing it for generalizing diverse rumor patterns. Extensive experiments are conducted on different benchmark datasets that demonstrated that the proposed DBi-GANet model significant outperforms other traditional deep learning techniques achieving high accuracy and robustness to linguistic variations.</p>