AI-Driven Privacy Audit Automation and Data Provenance Tracking in Large-Scale Systems

With the rapid development of mobile Internet, big data, supercomputing, sensor networks, brain science and other technologies, machine learning has entered a period of accelerated development to promote economic and social progress.However, its entire life cycle (data preprocessing, model training, model reasoning) is facing increasingly complex data security and privacy challenges, which is difficult for traditional protection technologies to cope with.This study proposes the following innovative solutions to address this issue: in the data preprocessing stage, a secure feature extraction scheme based on a single cloud server called SeiFS is developed, which integrates cryptographic primitives such as obfuscation circuits, unintentional transmission, and secret sharing to achieve end-to-end privacy protection; In the model training phase, two privacy preserving distributed training schemes (PEFL and PEFLimd) are designed, which combine additive homomorphic encryption and robust aggregation strategy to automatically filter poisoning gradients and protect gradient privacy[1-3]; In the model inference stage, an efficient convolution evaluation scheme based on homomorphic encryption (supporting large kernel and large step convolution) and a CryptoGT scheme for Transformer graph neural networks are proposed to reduce computational overhead and solve the "neighbor explosion" problem through fast convolution algorithms and hierarchical evaluation protocols.All schemes have passed formal security proofs and experimental verification.SeiFS achieves efficient security feature extraction in cloud computing scenarios, PEFL series schemes improve training efficiency while filtering poisoning attacks[4-5], CNN inference schemes significantly reduce homomorphic computational complexity, and CryptoGT effectively supports security evaluation of complex nonlinear functions, solving the scalability challenges of modern neural architectures and ensuring privacy as a whole.In the future, it is necessary to break through key issues such as lightweight and verifiable security feature extraction, universal security attack and defense system, non leakage model inference technology that preserves accuracy, and full lifecycle privacy protection for new generation models (such as big language models and video generation systems) [6][7].