Design of an Improved Model for Data Poisoning Detection Using AEAD-TL, GARNN, and FL-DPD

With the increased vulnerabilities of adversaries with respect to training data concerning machine learning models, the detection and prevention of data poisoning attacks in AI systems are now critical with strong mechanisms. Detect ion methods that have remained behind so far include fixed models for detection purposes but are not adaptable. This work addresses these limitations by providing novel and efficient detection and mitigation techniques against sophisticated poisoning attacks on various AI systems, with guarantees of accuracy and resilience under attack. We introduce five advanced models tailored to effectively counteract a wide range of poisoning attack types. First, the Adaptive Ensemble Anomaly Detection with Transfer Learning (AEAD-TL) enhances the detection process's accuracy through transfer learning of the ensemble of models for adapting evolving attack patterns. Those include Graph-Attack-Resistant Neural Networks (GARNN), which uses graph neural networks to detect attacks on the relationships of data in order to provide resistance to graph-based perturbations, followed by Autoencoder-Based Poisoning Gradient Monitoring (APGM), which monitors deviations in gradients during training in order to identify poisoned data points that would highly improve the detection accuracy with minimal overhead in terms of computations, and lastly, Federated Learning with Distributed Poisoning Defense (FL-DPD), which strengthens federated learning models by defending poisoned client updates using a mix of local and global anomaly detection mechanisms. The final model is the Resilient Hybrid Model with Transferable Outlier Detection, RHMT-OD, which integrates supervised learning and unsupervised learning to detect known as well as novel attacks, without requiring full retraining, while transferring learned patterns across datasets. Overall, these methods enhance the detection precision and reject false positives, while expediting the training procedure, thus ultimately enhancing the robustness of models in AI against poisoning attacks at minimal overheads for the system.