run inference ML Training Global Model Global Model Global Model Global Model Global Model Global Model Global Model Global Model Inference Inference Inference Inference Inference Inference
ML Global Model Global Model Global Model Global Model Global Model Global Model Global Model Global Model Inference Inference Inference Inference Inference Inference Model Aggregation
computation resource On-Device ML Inferencing ! responsive no network latency Federated Learning ! privacy preservation users don’t have to send raw data to server Recommendation User Interface Sensitive Data Treatment
computation resource ! responsive no network latency Federated Learning ! privacy preservation users don’t have to send raw data to server Recommendation User Interface Communication Info. On-Device ML Inferencing Server-side ML Federated Learning
computation resource ! responsive no network latency Federated Learning ! privacy preservation users don’t have to send raw data to server Recommendation User Interface Communication Info. Candidate Generation (1st Stage) Reranking (2nd Stage) On-Device ML Inferencing Server-side ML Federated Learning
- User features (estimated demographics, etc.) - Output (intermediate) - Item embeddings - User embeddings - Final output - Item candidates (per user cluster) Input Final Output Intermediate Output
- item embedding (for each candidate) - Output - score for each item - Client App. performs - inference, triggered by text input - training, triggered when device-idle (Personalized) Make use of intermediate output in 1st stage (Global) Input Output
upload without user ID - Differential Privacy (DP) mechanism Privacy Preservation Support multiple on-device ML instances - Separation of app. specific implementations from common FL functions
devices (Local DP) Support Differential Privacy (DP) mechanism Minimization of data collection - Federated Learning (collect ML models on behalf of raw data) - Model upload by randomly-sampled users without user id
the feasibility of FL - To-be: set a mature value that balances utility of FL and users’ privacy Seeking a privacy parameter ε Implementation of Local DP mechanisms with FL - Gaussian mechanism for local gradient (Local DP) - Averaging the aggregated local models (FL) - Local model upload without user id