Design of an Integrated Modern Approach to Detect and Prevent Data Poisoning Attacks in AI Systems: A Multi-Stage Defense Framework for Robust and Secure Learning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Adversarial threats like data poisoning attacks affect the training datasets and create biased, worse, or malicious AI models. Static heuristics and high false positive rates have impeded traditional defenses against such adaptive stealthy attacks. These defenses also do not generalize well in federated and non-IID settings. It is difficult to differentiate between poisoned and clean samples, and existing solutions even lack any adaptiveness to newly developed attack strategies; exhibit computational inefficiencies in process. This work presents a Modern Approach to Detect and Prevent Data Poisoning Attacks in AI Systems, coupled with a five-prong mechanism striving for data integrity and model resilience. Those mechanisms are Contrastive Adversarial Poisoning Detector (CAPD), the Federated Memory-Augmented Poison Filter (FMPF), Generative Adversarial Poison Purifier (GAPP), Graph Neural Network Consistency Analyzer (GNNC-A) and Adaptive Meta-Learning Defense System (AML-DS).All integrated, it achieves ≥ 96% poisoning detection accuracy, reduces attack success rates to ≤ 5%, and minimizes false positives to ≤ 2%.This work is scalable and adaptive and provides low computational overhead in security enhancement measures. By showcasing the high integrity of model security without excessively burdening the computer, this work significantly increases the robustness of AI systems in real-world applications.