Differential Privacy

Explanation

Differential Privacy checks that when a model is trained or evaluated on sensitive customer data, no single individual's records can be reverse-engineered from the model or its outputs - a core protection for bank data such as account balances, KYC details, or transaction histories used in Document Extraction and RAG pipelines. It is measured by the Differential Privacy Budget (epsilon): a parameter that quantifies how much the presence or absence of any one person's data can change the model's behaviour. A smaller epsilon means stronger privacy because individual contributions are mathematically masked by calibrated noise. To implement it, add a differential-privacy mechanism (for example, noise injection during training or query answering) at the data-processing and training stage of the pipeline, track the cumulative epsilon spent per dataset, and log the budget consumed for each training or evaluation run as auditable evidence. The threshold here is epsilon <= 1.0 for sensitive training and evaluation, which represents a strong privacy guarantee. Whenever the accumulated epsilon for a sensitive workload exceeds 1.0, the control should block or flag the run, require sign-off from the privacy office, and prevent further queries against that data until the budget is reset, aligning with RBI Data Privacy, NIST MEASURE 2.10, and ISO privacy-enhancing controls.