A Review of Research on Verifying the Integrity of Model Training Processes

Key words:  artificial intelligence  model training  training process integrity  integrity verification

Abstract:

The development of artificial intelligence technology and the demand for large-scale pre-trained models have led to the widespread use of third-party model training services, which are generally provided by cloud service providers. However, there is also a problem of the authenticity of the model training process -Specifically, whether the training process itself may be tampered with. Verifying the integrity of the model training process effectively ensures the reliability of the model training. While previous studies have mainly focused on the integrity of training data and privacy issues in model training for analysis, this paper primarily investigates the integrity of the entire process of model training. Integrity of model training is divided into three classes: computational correctness, algorithmic consistency and parameter update integrity. Then each category will be examined from the threat sources, Attack modes and Verification paths respectively. Finally, based on summarizing existing analysis results of model training integrity, it points out the direction for future research to verify the integrity of the model training process.