Research on the Risk Prevention and Control of Data Poisoning in the Age of Artificial Intelligence

The development of large artificial intelligence models relies on the integrity and accuracy of training data. Malicious contamination of training data introduces false information. This information becomes fixed as parameter bias during model training. Consequently, it affects judgment logic and model output. Therefore, data contamination has become a primary method for attacking the security of large models. Despite this threat, legal regulations concerning data poisoning lack systematic research. This paper employs a normative analysis method. It systematically reviews regulatory rules regarding data poisoning. These rules are found within Chinese criminal and data security laws. Current legislation primarily enforces general data security duties. It also relies heavily on accountability after an incident occurs. This approach makes early intervention difficult. Risks often evolve into model defects and cause damage before effective measures are taken. Therefore, the main contribution of this paper is moving beyond the current retrospective regulation model. We advocate establishing a risk control mechanism for the entire training process. This mechanism spans from prior review to continuous monitoring and subsequent accountability. Furthermore, it clearly divides the responsibilities among four main entities. These entities include data providers, model developers, service providers, and regulatory agencies. This mechanism helps resolve practical difficulties associated with data poisoning. These challenges include multiple participants, complex causal relationships, and unclear responsibilities. The proposed framework responds to the genuine needs of data security. It aligns with the governance goal of improving hierarchical and classified security supervision. Ultimately, it provides a feasible analytical framework for protecting training data.