How akumar2709/fides_asiaccs works
Based on our analysis of the file tree and README, this is a research/proof-of-concept implementation for result validation in Machine Learning-as-a-Service inference. It implements several distinct pipelines (training a service model, distilling a verification model, and training a detector/corrector) and also includes an attack testing script. The system is designed to work in a secure enclave environment using Gramine-SGX, with testing examples implemented using TFLite.
Main Components
• Service Model Training Component – Represented by the “training” folder (training.py, function_dispatcher.py, training.json). This module is responsible for training the service (inference) model.
• Distillation Verification Component – Represented by the “distillation” folder (distillation.py, function_dispatcher.py, distillation.json). It handles the distillation process that produces a streamlined (or “distilled”) version of the model for validation purposes.
• Detector/Corrector Training Component – Represented by the detector_corrector_training.py file. This script likely trains a mechanism to detect or correct results from the inference service.
• Attack Testing Module – Represented by attack_testing.py, which simulates attacks and tests the resilience (likely in low-cost result validation of the ML inference output).
Interactions and Data Flow
• The training component produces the service model.
• The distillation component takes input from the service model to train a verification model.
• The detector/corrector module operates by using its own training mechanisms and/or by taking inputs from the verification and service models.
• The attack testing module interacts with one or more of these components to simulate adversarial conditions.
• Also, indicate that the SGX secure environment is a deployment target for the inference service and verification models, highlighting the connection between the system’s output (e.g., TFLite models) and secure execution.
Architectural Patterns
• There is a pipeline or sequential flow:
• First, training the service model.
• Second, distillation for verification.
• Third, training a detector/corrector.
• Fourth, executing attack tests.
• This might be viewed as a multi-stage experimental pipeline which validates model performance and resilience under tested attack conditions.
• Loosely coupled modules – each folder or script is largely independent but is connected by data inputs/outputs and parameter configuration (through JSON files).
Technologies and External Dependencies
• That Gramine-SGX (and its documentation link) is used for SGX deployment.
• Usage of TFLite for testing – include a “TensorFlow Lite” node connected to the inference models.