25 min
Lesson 2 of 4
Technical Compliance Implementation
From the course AI Compliance & GDPR
Implement technical controls to achieve and demonstrate regulatory compliance in AI systems.
Data Protection by Design
Privacy-Preserving ML Techniques
# Differential Privacy with Opacus
from opacus import PrivacyEngine
model = YourModel()
optimizer = torch.optim.Adam(model.parameters())
dataloader = DataLoader(dataset, batch_size=64)
privacy_engine = PrivacyEngine()
model, optimizer, dataloader = privacy_engine.make_private_with_epsilon(
module=model,
optimizer=optimizer,
data_loader=dataloader,
epochs=10,
target_epsilon=1.0, # Privacy budget
target_delta=1e-5,
max_grad_norm=1.0
)
# Training with differential privacy guarantees
for batch in dataloader:
output = model(batch)
loss = criterion(output)
optimizer.zero_grad()
loss.backward()
optimizer.step()
print(f"Epsilon: {privacy_engine.get_epsilon(delta=1e-5)}")
Data Minimization Audit
# Identify feature importance to eliminate unnecessary PII
from sklearn.inspection import permutation_importance
result = permutation_importance(model, X_test, y_test, n_repeats=10)
# Create feature importance report
for i in result.importances_mean.argsort()[::-1]:
if result.importances_mean[i] - 2 * result.importances_std[i] > 0:
print(f"{feature_names[i]}: {result.importances_mean[i]:.3f}")
# Remove features with near-zero importance
# Especially PII features not contributing to accuracy
Right to Explanation
Model Explanation System
import shap
from flask import Flask, request, jsonify
app = Flask(__name__)
explainer = shap.Explainer(model, background_data)
@app.route('/predict-with-explanation', methods=['POST'])
def predict_with_explanation():
data = request.json['features']
# Get prediction
prediction = model.predict([data])[0]
# Generate explanation
shap_values = explainer([data])
# Format for human-readable explanation
feature_contributions = []
for i, (feature, value, shap_val) in enumerate(
zip(feature_names, data, shap_values.values[0])
):
if abs(shap_val) > 0.01: # Only significant factors
feature_contributions.append({
"factor": feature,
"value": value,
"impact": "increased" if shap_val > 0 else "decreased",
"magnitude": abs(float(shap_val))
})
return jsonify({
"decision": prediction,
"explanation": sorted(feature_contributions,
key=lambda x: -x["magnitude"])[:5],
"model_version": MODEL_VERSION,
"timestamp": datetime.utcnow().isoformat()
})
Consent Management
# Consent tracking for AI training data
class ConsentManager:
def get_valid_training_data(self, requested_data):
valid_records = []
for record in requested_data:
consent = self.db.get_consent(
user_id=record['user_id'],
purpose='ai_training'
)
if consent and consent.is_valid() and not consent.is_withdrawn():
valid_records.append(record)
return valid_records
def handle_deletion_request(self, user_id):
# Remove from all training datasets
self.remove_from_datasets(user_id)
# Retrain or use machine unlearning if feasible
self.queue_model_update(user_id)
Key Takeaways
- •Differential Privacy with Opacus provides mathematical privacy guarantees during training
- •SHAP-based explanation systems fulfill the right to explanation requirement
- •Consent management systems track and enforce data usage permissions
- •Data minimization audits using feature importance remove unnecessary PII
0 of 4 lessons marked complete
Progress is kept in this browser only.