This paper presents a bidirectional generative adversarial network (BiGAN) for detecting cyberattacks using the IoT23 database. The study addresses the challenge of data imbalance in intrusion detection systems (IDS) by leveraging deep learning (DL) and generative methods. The BiGAN model is trained and tested on the IoT23 dataset, which includes 21 instances of network traffic from three benign IoT devices and twenty samples of malware. The pre-processing phase involves removing redundant attributes, transforming categorical features, and normalizing the data. The BiGAN model consists of an encoder (E) and a generator (G), which are trained to invert each other to deceive the discriminator (D). The training process includes a reconstruction loss and an extra hint loss to enhance the model's ability to reconstruct input data. The proposed model achieved an accuracy of 98.8% and an F-score of 98.2%, outperforming conventional IDS models such as RF, SVM, LSTM, and AE. The results are validated through a 10-fold cross-validation, showing superior performance in detecting various types of cyberattacks. The study concludes by highlighting the advantages of using BiGANs in cybersecurity, including improved accuracy, reduced false positives, and adaptability to dynamic environments. Future work will focus on adapting the framework for federated learning and exploring adversarial attacks that could bypass generative DL-based IDS.This paper presents a bidirectional generative adversarial network (BiGAN) for detecting cyberattacks using the IoT23 database. The study addresses the challenge of data imbalance in intrusion detection systems (IDS) by leveraging deep learning (DL) and generative methods. The BiGAN model is trained and tested on the IoT23 dataset, which includes 21 instances of network traffic from three benign IoT devices and twenty samples of malware. The pre-processing phase involves removing redundant attributes, transforming categorical features, and normalizing the data. The BiGAN model consists of an encoder (E) and a generator (G), which are trained to invert each other to deceive the discriminator (D). The training process includes a reconstruction loss and an extra hint loss to enhance the model's ability to reconstruct input data. The proposed model achieved an accuracy of 98.8% and an F-score of 98.2%, outperforming conventional IDS models such as RF, SVM, LSTM, and AE. The results are validated through a 10-fold cross-validation, showing superior performance in detecting various types of cyberattacks. The study concludes by highlighting the advantages of using BiGANs in cybersecurity, including improved accuracy, reduced false positives, and adaptability to dynamic environments. Future work will focus on adapting the framework for federated learning and exploring adversarial attacks that could bypass generative DL-based IDS.