CNN 기반 ZKML 환경에서의 zk-SNARK 기반 데이터 무결성 검증 방법

초록

As Artificial Intelligence (AI) systems are increasingly applied in various decision-making environments, data integrity verification hasemerged as a critical challenge beyond ensuring the accuracy of inference results. In particular, it is necessary to verify that inference resultsare generated based on specific input data and that the data remains unaltered throughout the entire inference process. Existing studieshave mainly focused on model performance or the correctness of inference computations, while systematic mechanisms for verifying dataflow and temporal order within the inference pipeline have been limited. Although Zero-Knowledge Machine Learning (ZKML) approachesprovide strong guarantees for computational correctness, they have limitations in addressing temporal context anomalies such as replay andorder violations that may occur without bit level data modification. This paper proposes a method that integrates hash and timestampmechanisms into a Zero-Knowledge Succinct Non-Interactive Argument of Knowledge (zk-SNARK) based ZKML environment for data integrityverification in a single Convolutional Neural Network (CNN) model based AI inference setting. The proposed method verifies the data flowfrom preprocessing to inference output with temporal consistency and enables verification of the linkage between input data and inferenceresults without additional data disclosure or exposure of internal model information. Functional validation experiments conducted using normalinput data in a single CNN model based zk SNARK ZKML test environment confirm that data integrity is consistently maintained based onhash consistency. This study demonstrates the feasibility of verifying both data flow integrity and temporal consistency in AI inference pipelinesand provides a basis for data integrity verification in a single CNN model based AI inference environment.

키워드