pima-adisa-governance

Fast ADISA Demo & Machine Unlearning

This Streamlit application demonstrates a fast, lightweight Digit Classifier using an ensemble of specialized Neural Network models (FastCNN). It features instant predictions on handwritten images and an integrated lab to experiment with Machine Unlearning.

How Machine Unlearning is demonstrated in this app (SISA)

Machine Unlearning is the process of completely removing the influence of a specific training sample from a deployed model without the massive computational expense of retraining the model from scratch.

This app implements a SISA (Sharded, Isolated, Sliced, and Aggregated) paradigm:

1. Sharding: The training data is split into 5 entirely separate, disjoint chunks.
2. Isolation: We train 5 distinct “Expert” models, one on each chunk.
3. Tracking: The app maintains a dictionary (sample_to_expert) mapping exactly which data point went to which expert.
4. Targeted Unlearning: When the user requests to forget a sample (e.g., ID 42), the system queries the mapping to find the affected expert. It removes the sample from that expert’s local array and retrains only that specific expert.
5. Efficiency: The other 4 experts are completely untouched. This visualizes a massive speedup comparing the full ensemble retrain time vs the single-expert isolated retrain time. The Streamlit UI measures this execution time in realtime and generates a graph of the performance delta.