Governance of New Type of Data Security Risks in DeepSeek-like Generative Artificial Intelligence

DeepSeek类生成式人工智能的新型数据安全风险治理

Abstract

Data security risk governance in generative artificial intelligence (AI) is a significant current challenge. Existing data security governance systems for generative AI fail to establish effective, categorized governance mechanisms for different data security risks, resulting in a disconnect from the realities of evolving generative AI, particularly with models like DeepSeek. DeepSeek models, featuring multi-head latent attention mechanisms, hybrid expert models, and pure reinforcement learning, exhibit novel characteristics that alter or transform existing data security risks, leading to new data leakage, data deviation, and data deletion risks with varying operational logics. To address data leakage risks, both technical and managerial prevention mechanisms should be established. For data deviation risks, emphasis should be placed on fine-tuning both data input and output, different from manual oversight, and establishing algorithmic interpretation rights for data output results. For data deletion risks, a shift from absolute data deletion to relative deletion with dynamic verification mechanisms should be implemented.