Gradient Inversion Attacks: Impact Factors Analyses and Privacy Enhancement.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Gradient inversion attacks (GIAs) pose a significant threat to distributed learning by reconstructing private client data from shared model parameters.
• Existing privacy-preserving methods for distributed learning often compromise model usability or incur substantial computational overhead.
• The fundamental causes of data leakage in distributed learning remain underexplored, hindering effective defense strategies.

Purpose of the Study:

• Investigate the underlying reasons for the success of current GIAs.
• Analyze model robustness against GIAs throughout the training process.
• Examine the influence of different model architectures on GIA effectiveness.

Main Methods:

• Exploratory analysis of GIA vulnerabilities in distributed learning.
• Empirical investigation of model robustness variations during training.
• Evaluation of model structure impact on attack performance.
• Development of a plug-and-play defense method using vicinal data augmentation.

Main Results:

• Identified key factors contributing to successful GIAs.
• Demonstrated variations in model robustness during distributed training.
• Quantified the impact of model architecture on attack success.
• Empirical validation of the proposed defense method's effectiveness.

Conclusions:

• The proposed plug-and-play defense method effectively enhances privacy in distributed learning.
• This approach augments training data with a vicinal distribution, ensuring basic privacy levels.
• The method maintains global model usability while defending against gradient inversion attacks.