Online Domain Adaptation Framework for Streaming Data Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing online learning frameworks struggle to preserve user privacy due to the distribution shift from public data to private user data, and they fail to effectively implement the Right to Be Forgotten by not deleting user data after use.
Innovation Solution
An online domain adaptation framework based on cross-domain bootstrapping is introduced, where target domain streaming data is deleted immediately after adaptation. This framework increases data diversity by bootstrapping the source domain and uses independent machine learning models to preserve differences, integrating their knowledge through pseudo-labels to maintain divergence without sharing weights.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If user data is retained for training machine learning models, then model accuracy is improved, but user privacy is compromised
Solution Approach 1:
The system segments the training process into two distinct phases: offline pre-training on public source domain data, and online adaptation on private target domain data. This segmentation allows the model to learn general patterns from public data without retaining private user data, thereby maintaining accuracy while protecting privacy.
Solution Approach 2:
The system implements discarding of target domain data immediately after it is used for online adaptation. The knowledge gained from this data is recovered and transferred to the source domain through the domain adaptation mechanism, eliminating the need to store private user data while preserving its informational value.
2Object-affected harmful factors
If user data is deleted after use, then privacy protection is achieved, but the Right to Be Forgotten cannot be implemented
Solution Approach 1:
The system performs preliminary domain adaptation training using target domain data before any deletion occurs. The offline pre-training phase prepares the model to effectively adapt to new domains, ensuring that when target data is deleted, the adaptation knowledge has already been transferred to the source domain representations.
Solution Approach 2:
The system creates a copy of the adaptation knowledge by transferring learned patterns from the target domain back to the source domain. This copying mechanism ensures that the essential information is preserved in the source domain model, allowing complete deletion of target data while maintaining the Right to Be Forgotten.
3Adaptability or versatility
If multiple independent machine learning models are used to preserve data differences, then model diversity is improved, but system complexity increases
Solution Approach 1:
The system merges multiple independent learner models into a unified domain adaptation framework. The learners are coordinated through a shared source domain representation and a domain adaptation module that integrates their predictions, reducing overall system complexity while preserving model diversity and adaptability.
Solution Approach 2:
The source domain model serves multiple functions: it acts as a shared representation for all learners, provides a basis for domain adaptation, and enables knowledge transfer back to the source domain. This multi-functionality reduces the need for separate components, simplifying the system while maintaining diversity.
Data Source
AI summary
Embodiments described herein provide an online domain adaptation framework based on cross-domain bootstrapping for online domain adaptation, in which the target domain streaming data is deleted immediately after adapted. At each online query, the data diversity is increased across domains by bootstrapping the source domain to form diverse combinations with the current target query. To fully take advantage of the valuable discrepancies among the diverse combinations, a set of independent learners are trained to preserve the differences. The knowledge of the learners is then integrated by exchanging their predicted pseudo-labels on the current target query to co-supervise the learning on the target domain, but without sharing the weights to maintain the learners' divergence.