METHODS FOR COUNTERING DATA POISONING DURING THE FINE-TUNING PHASE OF NEURAL NETWORKS
Background. The research investigates how data poisoning attacks during fine-tuning stages make modern neural networks extremely vulnerable to attacks. The industry now focuses on domain-specific datasets, which attackers use to create unequal market advantages through their ability to hide backdoors with minimal changes to data. The current protection methods fail to provide proper security because they cause major memory loss of essential information, while attackers can bypass their protection mechanisms by using complex semantic attacks against systems which operate in decentralized federated learning environments. The research develops a multi-loop protection system which protects the latent feature space through mathematical validation. The proposed method combines neural network retraining with reverse analysis of synaptic weight changes to detect compromised neurons at a high level of precision, which does not affect the model's performance when working with unaltered data. The research bases its theoretical structure on the concept which protects distributed infrastructures through multiple layers of defense. The system uses local stochastic filtering operations which run on client devices through DP-SGD differential privacy mechanisms before servers perform semantic auditing with explainable artificial intelligence (XAI) for protection. This is then followed by smart contract-based cryptographic verification of update processes. The experimental data shows the results of the study. The developed architecture shows strong performance through empirical testing, which used benchmark architectures that included convolutional neural networks (CNNs) and large language models (LLMs) for critical application systems. The proposed multi-loop filter implementation enables the Attack Success Rate to achieve a zero limit, while it maintains the model's ability to predict correctly for legitimate data. Research conclusions. The results obtained from a robust theoretical and algorithmic foundation for the transition from reactive anomaly detection methods to proactive topological protection of intelligent systems. The developed methodology guarantees the structural integrity of neural network models when transfer learning is applied in high-risk domains.