AWS Clean Rooms ML
A privacy-preserving ML collaboration feature that enables multiple organizations to jointly train and run inference on machine learning models without sharing raw data
Overview
AWS Clean Rooms ML is an extension of AWS Clean Rooms that enables multiple organizations to jointly execute machine learning model training and inference without directly sharing raw data. It allows data collaboration partners such as advertisers and publishers, or retailers and manufacturers, to build lookalike audiences and conversion prediction models while preserving privacy. It leverages differential privacy and secure computation techniques to mathematically control the risk of individual record identification. It makes cross-organization ML utilization, previously difficult due to legal and technical barriers, easily accessible as a managed service.
Lookalike Audience Generation with Lookalike Modeling
The representative use case for Clean Rooms ML is Lookalike Modeling (lookalike audience generation). It matches an advertiser's high-value customer list (seed data) against a publisher's user attribute data to generate new target audiences with characteristics similar to the seed customers. This processing executes within the Clean Rooms collaboration environment; the advertiser cannot see the publisher's individual user data, and the publisher cannot see the details of the advertiser's customer list. The generated lookalike audience is output as segments with similarity scores, which can be linked to ad delivery platforms to improve targeting precision. Model training is automated, allowing execution from the console in just a few clicks without ML expertise.
Technical Mechanisms for Privacy Protection
Clean Rooms ML's privacy protection consists of multiple technology layers. First, collaboration analysis rules restrict the scope and aggregation level of data each participant can access. In ML processing, differential privacy noise injection makes it mathematically difficult to reverse-engineer individual records from model outputs. The privacy budget concept manages the accumulation of privacy risk through cumulative queries, automatically blocking queries that exceed the budget. Related books on privacy protection (Amazon) cover the theory of differential privacy. Secure computation over encrypted data ensures data is never exposed in plaintext during processing. Audit logs are recorded in CloudTrail, enabling tracking of who executed what analysis and when.
Collaboration Design and Practical Considerations
Utilizing Clean Rooms ML requires first establishing collaboration agreements between partners. Data usage purposes, analysis rules (minimum aggregation units, permissible output metrics), and privacy budget allocation must be determined in advance and reflected in Clean Rooms collaboration settings. In practice, legal department review of data usage agreements and technical team design of analysis rules typically proceed in parallel. The cost structure is determined by data volume processed within the collaboration and ML job execution time. For Lookalike Modeling, seed data size and matching target data scale directly impact pricing, so an approach that improves seed quality while reducing quantity (limiting to the top 10% of high-value customers, for example) offers superior cost efficiency. Model retraining frequency is set weekly to monthly depending on audience change velocity, balancing freshness against cost.