Sparse ANNS: Hardware Aware Agentic Inference for Real Time Fraud Detection via Sparse Memory Hierarchies

Bidhan Biswas; MD Towheduzzaman; Ali Ashraf Tanvir

doi:10.36948/ijfmr.2025.v07i03.78516

Sparse ANNS: Hardware Aware Agentic Inference for Real Time Fraud Detection via Sparse Memory Hierarchies

Author(s)	Mr. Bidhan Biswas, Mr. MD Towheduzzaman, Mr. Ali Ashraf Tanvir
Country	United States
Abstract	We propose Sparse-ANNS, a hardware-aware inference engine for Detector Agents in Multi-Agent Anomaly Detection Systems (MAS-ADS), designed to achieve real-time anomaly detection with sub-millisecond latency. The increasing demand for low-latency fraud investigation in high-velocity transactional data streams necessitates a paradigm shift from dense to sparse computation, yet existing systems often fail to balance efficiency and accuracy. The proposed method integrates sparse tensor computation with Approximate Nearest Neighbor Search (ANNS), reformulating the inference pipeline into three novel modules: a sparse feature embedding layer with pruned attention, a hardware-optimized ANNS graph with memory hierarchy-aware partitioning, and a dynamic thresholding unit calibrated via graph regularization. These components collectively address the inefficiencies of traditional dense processing while preserving detection performance. Moreover, the system employs non-uniform quantization for input compression and federated meta-learning for adaptive ensemble scoring, enabling seamless integration with conventional MAS-ADS modules. Implemented with state-of-the-art technologies such as TensorRT-LLM and FAISS-IVFPQ, Sparse-ANNS demonstrates a 92% reduction in 99th-percentile inference latency compared to baseline systems, maintaining an AUROC above 0.95 on real-world fraud datasets. The significance of this work lies in its unified approach to sparse computation and hardware-aware optimization, which bridges the gap between theoretical advances in ANNS and practical constraints of real-time agentic systems.
Keywords	Sparse Computation, Approximate Nearest Neighbor Search (ANNS), Multi-Agent Anomaly Detection Systems (MAS-ADS), Hardware-Aware Inference, Real-Time Fraud Detection, Sparse Tensor Processing, Federated Meta-Learning, Graph Regularization, Low-Latency AI, FAISS-IVFPQ, TensorRT-LLM, Dynamic Thresholding, Edge AI Optimization, Adaptive Ensemble Scoring, High-Velocity Transaction Streams
Field	Computer > Network / Security
Published In	Volume 7, Issue 3, May-June 2025
Published On	2025-05-09
DOI	https://doi.org/10.36948/ijfmr.2025.v07i03.78516

View / Download PDF File

About IJFMR Fees & Payment Current Issue Publication Archive	Submit Research Paper Track Submission Status Publication Guidelines Publication Ethics Peer Review & Plagiarism	Join as a Reviewer Editors & Reviewers Reviewer Referral Program Get Reviewer Membership Certi.	Website/Journal Policies Usage Policy Content Policies Privacy Policy

Contact Us		+91-9687-828-838	editor@ijfmr.com

International Journal For Multidisciplinary Research

A Widely Indexed Open Access Peer Reviewed Multidisciplinary Bi-monthly Scholarly International Journal

Sparse ANNS: Hardware Aware Agentic Inference for Real Time Fraud Detection via Sparse Memory Hierarchies

Share this