Explainable AI: Feature Attribution & Privacy Simulator

Model Inputs (Features)

Income score0.60

Credit history0.55

Debt ratio0.40

Recent inquiries0.30

Privacy vs. Explainability

Privacy budget ε 2.0

Model output

–

Decision

–

Explanation drift

–

Top driver

–

How it works

The dark cube is a "black box" credit-approval model: it takes four input features and turns them into a single approve/deny decision. Instead of trusting the box blindly, an explainability layer (LIME/SHAP-style) computes each feature's attribution — how much that feature pushed the output toward approve (green) or deny (red) relative to an average applicant. The four beams and floating bars visualize that attribution: thicker/taller means more influence.

logit = bias + Σ wᵢ·xᵢ
attribution_i = wᵢ·(xᵢ − baseline)   // Shapley value, linear case
p = sigmoid(logit),  decision = p > 0.5

Feature sliders — the applicant's input values; moving one instantly recomputes the model's output and every feature's attribution.
Privacy budget ε — controls differential-privacy noise added to the model and its explanations before they're shown; a small ε (strong privacy) injects large random noise, while a large ε (weak privacy) leaves attributions almost exact.
DP noise toggle — turn the noise off to see the "ground truth" attribution the model actually used, or on to see what a privacy-preserving release of the same explanation looks like.
Explanation drift — the average size of the noise injected into the attributions; high drift means the explanation you're shown may not match the real reasoning, a core interpretability/safety trade-off.

Real-world relevance: regulated ML systems (credit, hiring, medical triage) increasingly must explain individual decisions, but publishing exact per-feature attributions can leak information about the training data — differential privacy trades some explanation fidelity for a mathematical guarantee that no single record can be reverse-engineered from it.

Pushes toward Approve Pushes toward Deny Black-box model core

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.