The exponential expansion of consumer financial services has led to massive volumes of structured transaction and account data being migrated to public cloud data lakes for high-throughput distributed analytics. However, financial datasets intrinsically contain sensitive Personally Identifiable Information (PII) subject to stringent regulatory frameworks including the General Data Protection Regulation (GDPR) and the Indian Digital Personal Data Protection (DPDP) Act of 2023. Conventional privacy solutions exhibit a debilitating trilemma: cryptographic methods such as Fully Homomorphic Encryption (FHE) impose prohibitive computational latency; static dictionary-based anonymization collapses under schema drift and lexical variations; and unconstrained machine learning classifiers suffer from train/serve distribution shifts where subtle format discrepancies precipitate catastrophic privacy leaks.

In this paper, we propose a scalable, cloud-native privacy preserving big data analytics framework anchored by an Adaptive Dual-Tier Privacy Engine. Our architecture couples an adaptive multi-stage gradient boosted decision tree (XGBoost) classifier— trained on a 25-dimensional structural, keyword, sample-level, and regular expression feature space—with deterministic safety-net guardrails (Force-Protect and Force-Keep). This hybrid design guarantees zero exposure of mission-critical identifiers while completely insulating downstream business analytical features against corruption. Masked datasets are securely persisted to AWS S3 and processed through distributed Apache Spark (PySpark) clusters via the S3A connector across four analytical dimensions: descriptive statistics, bivariate fraud detection with two-factor authentication (2FA) correlation, credit risk profiling, and multi tier customer segmentation. Validated against a high-cardinality 100,000-record banking benchmark across 57 attributes, our system demonstrates 100% mitigation of critical PII leakage, overcomes real-world distribution skew on complex identifier formats, preserves analytical utility with zero loss of statistical fidelity, and maintains sub-minute distributed pipeline throughput.

Index Terms—Privacy-Preserving Big Data Analytics, Cloud Computing, Apache Spark, XGBoost, Personally Identifiable Information (PII), Hybrid Guardrails, Train/Serve Skew, Financial Data Governance.