Dynamic Adversarial Data Collection for MNIST Model in HuggingFace Spaces

AI Impact Summary

Dynamic Adversarial Data Collection (DADC) replaces static benchmarks with a human-in-the-loop feedback loop to repeatedly collect adversarial examples and retrain the MNIST model. This approach improves robustness and generalization to diverse handwriting, but requires building data collection, labeling, and retraining pipelines, and integrating with user-facing tools like HuggingFace Spaces and Gradio. Operational considerations include data governance for user-contributed samples, latency of retraining cycles, and ensuring flagging and dataset curation maintain data quality.