BlazingAML: High-Throughput Anti-Money Laundering (AML) via Multi-Stage Graph Mining
The system uses a novel compiler to transform high-level pattern descriptions into optimized CPU/GPU code.
A team of researchers including Haojie Ye, Arjun Laxman, and Krisztian Flautner has published a paper on BlazingAML, a high-throughput system designed to revolutionize anti-money laundering (AML) detection. The system directly addresses core industry challenges: excessive false positives and an inability to adapt to sophisticated, multi-stage laundering schemes that exploit modern financial networks. While graph analytics and AI are promising tools, they traditionally struggle with the inherent 'fuzziness' of laundering patterns, which exhibit significant structural and temporal variations. BlazingAML's novel contribution is a multi-stage framework that decomposes complex schemes into logical stages connected by graph operations. This allows diverse, fuzzy patterns to be expressed using a unified set of primitives, moving beyond the need for analysts to manually enumerate every pattern variant—a process that is both error-prone and computationally expensive.
The second major innovation is a domain-specific compiler that transforms these high-level, analyst-friendly pattern descriptions directly into highly optimized code for both CPU and GPU back-ends. This compiler applies sophisticated parallelization and optimization techniques automatically, completely eliminating the need for financial crime analysts to be experts in manual parallel programming. The performance results are staggering. When evaluated on real-world IBM AML datasets, BlazingAML achieved the same critical detection accuracy (F1 score) as current state-of-the-art methods. However, it did so while delivering a 210x higher speedup on CPU systems and a 333x higher speedup on GPU systems, demonstrating superior scalability. This monumental increase in throughput means financial institutions can analyze vastly larger transaction graphs in near real-time, potentially identifying complex, cross-border laundering rings that were previously undetectable due to computational constraints.
- Achieves 210x speedup on CPU and 333x on GPU while matching state-of-the-art detection accuracy (F1 score) on IBM datasets.
- Introduces a multi-stage graph mining framework to model 'fuzzy' laundering patterns with structural and temporal variations.
- Features a domain-specific compiler that auto-generates optimized parallel code, removing manual programming needs for analysts.
Why It Matters
Enables financial institutions to detect sophisticated, multi-stage money laundering schemes in real-time, drastically reducing false positives and operational costs.