Study of Privacy Preservation Techniques for Deep Learning
Abstract
Machine Learning (ML) as-a-service (MLaaS) has brought much convenience to our daily lives. However, these MLaaS are often offered through cloud computing services which raises the potential risk of privacy leakage when personal data were used in the model development. We propose to implement Privacy-Preserving Machine Learning (PPML) through data transformation, where the data is first transformed through nonlinear lossy compression mapping mechanism before sending to the cloud to have the ML service. The transformed data is not reversible and thus, the data privacy could be preserved. Moreover, the most important information for ML could be retained for the ML service in the cloud. The nonlinear lossy compression mapping mechanism can be evaluated by how difficult the adversary can perform reconstruction attack based on the nonlinearly compressed data, while as how well the MLaaS can provide its service accordingly. This in turn can be formulated as an adversarial learning problem resembling Generative Adversarial Network. In this proposal we focus on (1) The implementation of PPML to various applications such as panorama image and video data; (2) The application of secure multi-party computation (SMPC) to the training of neural network as well as its speedup for practical large-scale datasets, so as to further enhance privacy protection on the training data.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 27, 2024
- Accession Number
- AD1229984
Entities
People
- Pei-Yuan Wu
Organizations
- National Taiwan University