Requirements
Functional Requirements
- Accept pre-trained image classification models (e.g., ResNet50, EfficientNet, Vision Transformer (ViT))
- Support dataset input (local folders / standard datasets like CIFAR-10, COCO)
- Allow user to specify:
- Target label (e.g., “dog”)
- Number of test samples
- Execute multiple testing strategies:
- Random Testing
- Stress Testing
- AI-Guided Testing
- Feature-Based Testing
- Generate:
- Metrics (Accuracy, Precision, Recall, F1-score)
- Failure summaries
- Visual comparisons (before vs after transformations)
Non-Functional Requirements
- Scalability (handle large datasets)
- Extensibility (add new testing strategies)
- Interpretability (clear explanations via LLM)
- Usability (simple UI + clear reports)
Prototype Design
System Architecture
- Frontend (React + Vite)
- User input (model, dataset, label)
- Visualization (graphs, results, images)
- Backend (Node.js + Express)
- API endpoints:
- /model-report
- /random-test
- /stress-test
- /feature-test
- API endpoints:
- ML Execution Layer (Python)
- Model inference (PyTorch / TensorFlow)
- Image transformations
- LLM Integration
- Summary generation
- Failure analysis
- Feature extraction
Data Flow
- User inputs model + dataset
- Backend initializes testing pipeline
- Python layer executes tests
- Results → sent to LLM
- LLM generates insights
- UI displays final report
Completed Work
- Implemented:
- Random Testing pipeline
- Basic Stress Testing (background, occlusion)
- Model summary using LLM
- Integrated:
- Pre-trained models (ResNet, VGG)
Observations
- Accuracy alone is insufficient
- Models fail under:
- Occlusion
- Background changes
- Confidence scores do not reflect robustness
Challenges
- Mapping dataset labels to model classes
- Generating meaningful failure explanations
- Handling API limits (LLM cost)
Final Design (Detailed)
Core Components
1. Model & Dataset Analyzer
- Extract:
- Architecture details
- Training assumptions
- LLM generates:
- Model summary
- Suggested testing strategy
2. Random Testing Module
- Input: label + dataset
- Process:
- Sample images randomly
- Run inference
- Output:
- Pass/Fail counts
- Accuracy metrics
3. Stress Testing Module
- Apply transformations:
- Background change
- Environment shift
- Occlusion
- Multi-object addition
- Compare:
- Before prediction
- After prediction
4. AI-Guided Testing Module
- LLM generates:
- Edge cases
- Rare scenarios
- Examples:
- Dog in unusual pose
- Low-light conditions
5. Feature-Based Testing Module
- LLM extracts:
- Visual features (e.g., “fur texture”, “ear shape”)
- Testing:
- Evaluate model on feature combinations
6. Failure Analysis Engine
- Input:
- Failed test cases
- LLM identifies:
- Patterns
- Root causes
Examples:
- “Model fails when object is partially occluded”
- “Confusion between visually similar classes”
7. Reporting Module
- Generates:
- Summary report
- Graphs (performance trends)
- Recommendations
A. Introduction for reporting:
Problem: Lack of robust testing frameworks
Motivation: Hidden failures in ML models
B. Related Work:
Uncertainty-based testing (FAST)
Traditional test selection methods
Limitations: No semantic understanding
No interpretability
C. Proposed Framework:
Unified testing system
Integration of:LLM
Multiple testing strategies
D. Methodology:
Testing Types
Random Testing → baseline
Stress Testing → robustness
Feature-Based Testing → interpretability
AI-Guided Testing → exploration
E. Experimental Setup:
Models: ResNet50, EfficientNet, ViT
Datasets:CIFAR-10, COCO (subset)
F. Evaluation Metrics:
Accuracy, Precision / Recall, F1-score
Failure Rate:
APFD (test prioritization)
G. Results & Analysis:
Stress testing reveals hidden failures
Feature-based testing improves interpretability
AI-guided testing finds edge cases
H. Contributions:
Unified AI-assisted testing framework
Integration of LLM for explanation
Test generation
Feature-level testing approach