Data poisoning attack classification with ConvNeXT-GAN

Deep Learning (DL) models are highly susceptible to data poisoning attacks, which can degrade model reliability and compromise trust in real-world applications. To address this challenge, this paper proposes a novel ConvNeXT-GAN architecture specifically designed to enhance resilience against data poisoning. The proposed framework uniquely integrates ConvNeXT residual blocks within the Generative Adversarial Network (GAN) discriminator, enabling deeper and more stable feature extraction that strengthens the model’s ability to mitigate the effect of adversarial flipped data. The robustness of ConvNeXT-GAN is evaluated under both random and distance-based label-flipping attacks, and compared against traditional and recent state-of-the-art models, including GAN, Gradient Penalty GAN (GPGAN), Federated GAN (FGAN), Long short term memory (LSTM) network, Autoencoder, Residual Network (ResNet), Certified Robustness via Finite Aggregation (CRFA), Robust Foundation Models Against Clean-Label Poisoning (RFM-CLP), Differentially Private Self-Normalizing Neural Networks (DiPSeN), StyleGAN2, BigGAN, and Vision Transformers (ViT). Experimental results show that ConvNeXT-GAN achieves superior accuracy of 0.952, 0.9703, 0.964, and 0.971 on poisoned datasets, significantly outperforming existing approaches. These outcomes confirm that the ConvNeXT-GAN architecture offers a robust and effective defense mechanism against data poisoning attacks while maintaining high predictive performance.