Research Paper & Specification

SATE AI: Systematic Fault Injection Framework for On-Device AI Models in Flutter

Antigravity & SATE AI Research Team

Download PDF Paper Download LaTeX Source (.tex)

Abstract

On-device Artificial Intelligence (AI) models operating directly on mobile and edge devices face unpredictable runtime constraints including memory pressure, thermal throttling, and unvalidated user inputs. Existing mobile testing paradigms lack specialized fault injection mechanisms tailored for AI inference runtimes. In this work, we present SATE AI, a fault injection and evaluation framework for Flutter applications that systematically evaluates model robustness across seven key failure modes.

1. Software Architecture

SATE AI enforces a decoupled architecture isolating fault injectors, model adapters, and orchestration runners. This allows model runtimes (such as ONNX Runtime and TensorFlow Lite) to be evaluated under simulated stress scenarios without modifying the core model execution engine.

SATE AI Architecture
Figure 1 – SATE AI Architecture showing core components, adapters, and fault injectors.

The framework features seven core fault injectors capable of targeting memory pressure, thermal degradation, quantization drift, latency spikes, input corruption, model swapping, and confidence validation.

Fault Injectors Overview
Figure 3 – The seven fault injectors and their failure modes.

2. Fault Injection Workflow

During stress execution, the StressRunner coordinates sequential or grouped fault injections against an AIModelAdapter instance. Each injector alters internal model state, monitors inference outputs, measures latency overhead, and asserts error handling compliance.

Fault Injection Workflow
Figure 2 – The fault injection workflow: from model adapter to stress report.

3. Implementation & Example Usage

SATE AI provides a clean programmatic Dart API for developers to wrap custom on-device runtimes, define targeted injection thresholds, and process structured StressReport outputs in JSON or Markdown format.

Example Usage Code Flow
Figure 4 – Code example and its execution flow.

4. Availability & Downloads

SATE AI is open-source and published on pub.dev. Full LaTeX sources, bibliography data, and research artifacts are accessible via the repository.