International Journal of Scientific Research and Engineering Development

Data scarcity and stringent privacy regulations—including GDPR, HIPAA, and India's Digital Personal Data Protection (DPDP) Act—critically impede the development of data-driven AI systems. Existing synthetic data generators are either restricted to a single modality, computationally prohibitive, or offer inadequate privacy guarantees, making them unsuitable for practical multi-domain deployment. We present PrivSynth, a unified, lightweight framework that simultaneously generates high-fidelity synthetic data across three modalities: structured tabular data via Conditional Tabular GAN (CTGAN), sequential time-series data via TimeGAN, and natural-language text via GPT-2 with Low-Rank Adaptation (LoRA) fine-tuning. Privacy is enforced through a novel dual-layer mechanism that combines DP-SGD training (providing formal (ε, δ)-differential privacy guarantees) with nearest-neighbour post-generation filtering (blocking record memorisation at inference time). An integrated AutoML module jointly optimises generator hyperparameters and the privacy budget ε without expert intervention. Domain-aware adapters enable zero-shot domain switching, and INT8 quantisation ensures edge deployability with <2% utility loss. Evaluated on three public benchmarks—UCI Adult, UCI Electricity Load, and AG News—PrivSynth achieves a downstream classification accuracy of 0.868 ± 0.009, AUC of 0.881 ± 0.007, and F1-score of 0.854 ± 0.011 under a privacy budget ε = 1.8, outperforming four competitive baselines while attaining a membership-inference attack success rate of 0.513—near the theoretical random-guessing floor of 0.500.

Shifa Bilal Tamboli, Simeen Phiroj Mulani, Arman Tajuddin Shiakh,"PrivSynth: A Unified Privacy-Preserving Synthetic Data Framework with Dual-Layer Differential Privacy, Auto ML, and Edge Deployment" International Journal of Scientific Research and Engineering Development, V9(2): Page(2484-2489) Mar-Apr 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.