MUMBAI, India, Aug. 24 -- Intellectual Property India has published a patent application (202641028024 A) filed by Garlapati Veera Venkata Sai Raviteja on March 10, 2026, for System And Method For Adaptive Anomaly Detection Using Dual-Agent Adversarial Co-Evolution.
Inventor includes Raviteja Garlapati V V S.
The application for the patent was published on August 21, 2026, under issue no. 34/2026.
Abstract: A dual-agent adversarial learning system and method for financial fraud detection is disclosed. The system comprises Agent CRIMSON, a generative attacker agent that synthesises fraudulent transaction samples through a blend-ratio interpolation mechanism continuously interpolating between real fraud and legitimate transaction statistical distributions, and Agent COBALT, an adaptive defender agent that retrains iteratively on complete historical data augmented with adversarially generated samples. A dual-batch architecture separates evasion-optimized samples (for defender training on hard patterns) from unbiased probe samples (for measuring defender detection capability). A three-signal adaptive strategy module adjusts attacker parameters — blend ratio and noise scale — based on simultaneous inputs from probe detection rate, evasion rate, and a drift-score guardrail that constrains synthetic sample realism. Reproducibility validation across three independent random seeds confirms consistent performance improvement: mean Accuracy improvement +0.82% (±0.07%), mean ROC-AUC improvement +0.37% (±0.01%), mean F1 improvement +0.66% (±0.06%), with ROC-AUC standard deviation of 0.00% indicating near-deterministic discrimination capability. In the primary experimental embodiment on 50,000 financial transactions, test accuracy reached 94.51%, ROC-AUC 98.70%, and F1 95.26%, with false negatives reduced by 20.6% (354 to 281). The adversarial effectiveness score reached 0.994 under a four-component evaluation metric. The invention addresses the fundamental limitation of static fraud detection systems by proactively generating and learning from fraud patterns that have not yet occurred in historical data.
Disclaimer: Curated by HT Syndication.
