MUMBAI, India, Oct. 12 -- Intellectual Property India has published a patent application (202641119812 A) filed by Rns Institute Of Technology, Bengaluru; Bhuvan Srncs; Adarsh Nrncs; Aryan Rajurncs; Ms. Vinuta A Patil Assistant Professor; Ms. Vanishri Sataraddiassistant Professor; and Parthiv Chandrurncs on October 07, 2026, for Privacy-Preserving Federated Graph Neural Network For Multi-Strategy Income Tax Optimization.
Inventors include Ms. Vanishri Sataraddiassistant Professor; Ms. Vinuta A Patil Assistant Professor; Aryan Rajurncs; Bhuvan Srncs; Adarsh Nrncs; and Parthiv Chandrurncs.
The application for the patent was published on October 09, 2026, under issue no. 41/2026.
Abstract: The present invention relates to TaxMind-FL, a privacy-preserving federated graph neural network system for multi-strategy income tax optimization that integrates heterogeneous financial graph construction, on-device Graph Attention Network processing, federated model aggregation, adaptive differential privacy, Pareto-optimal tax-route generation, and explainable deduction-level attribution within a unified platform. The proposed system comprises a client-side financial modeling unit equipped with a heterogeneous financial graph constructor, two-layer multi-head Graph Attention Network, gradient clipping and Gaussian noise injection module, and a Pareto-optimal strategy generator. The financial graph constructor encodes each user's income, deduction, and tax-regime information as nodes and edges, while the attention network learns which deductions are most relevant to a given taxpayer. A local training routine computes gradient updates on-device, and a lightweight FedAvg-based federated aggregation mechanism deployed on a central server combines these updates into a global model without accessing raw financial data. The current classification framework categorizes each taxpayer's optimal strategy into two regime categories: Old Regime and New Regime. Graph-based feature information and attention-weighted neighbor aggregation are combined through the Graph Attention Network to determine the final regime classification. The resulting classification, together with per-deduction feature attributions, is provided to the strategy engine, which generates a ranked set of Pareto-optimal tax-saving routes for the corresponding regime. Each client device further acts as an edge-processing and privacy-preserving gateway for transmitting clipped and noise-perturbed gradient updates, rather than raw financial data, to a centralized aggregation server. The server aggregates the received updates using FedAvg and provides the refined global model for on-device generation of tax-saving strategies, without storing or accessing any user's underlying financial records. The Pareto optimizer evaluates feasible deduction combinations across both tax regimes and returns a non-dominated set of routes ranked by net savings and compliance burden. The SHAP explainer decomposes each recommendation into deduction-level attributions that sum to the difference between the model output and its expected value, providing a legally auditable rationale for every suggested strategy. A privacy accountant tracks the cumulative differential-privacy budget across the annual training cycle using the Moments Accountant, ensuring the formal guarantee is never exceeded. The integrated architecture provides a unified approach to privacy-preserving tax classification, federated model refinement, multi-strategy optimization, and auditable explanation, while providing a modular platform capable of future expansion.
Disclaimer: Curated by HT Syndication.
