Transaction Constraint Tuning Through Adversarial Reinforcement Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current transaction monitoring systems rely on deterministic rules to detect suspicious activity, but the evaluation of these rules is inadequate as they do not accurately measure the value of a scenario or rule in identifying suspicious activity, often leading to unclear attribution of effective cases and inefficiencies in rule improvement.
Innovation Solution
Implement an adversarial reinforcement learning agent to evaluate transaction monitoring systems by measuring how often products are used to evade scenarios without triggering alerts, and automatically modify transaction constraints to reduce illicit use, using an RL agent to quantify the effectiveness of monitoring systems and strengthen weak points.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deterministic rules are used to detect suspicious activity, then known forms of suspicious activity can be detected, but the system cannot accurately measure the value of scenarios or rules in identifying suspicious activity
Solution Approach 1:
Instead of having the monitoring system detect suspicious activity directly, an adversarial RL agent is introduced to attempt evasion of the monitoring system. This inverted approach allows indirect measurement of scenario effectiveness by observing which scenarios the agent successfully bypasses, thereby providing accurate measurement of scenario value without direct complexity in the evaluation mechanism
Solution Approach 2:
An adversarial reinforcement learning agent serves as an intermediary between the monitoring system and the evaluation process. This agent mediates the measurement of scenario effectiveness by attempting to evade detection, allowing the system to indirectly assess which scenarios are most effective at identifying suspicious activity while maintaining manageable complexity
2Reliability
If multiple alerts lead to a single case of suspicious activity, then detection coverage is improved, but it becomes unclear which alert or scenario should be credited for the effective case
Solution Approach 1:
The evaluation process is segmented by having the adversarial agent attempt evasion through specific products or channels. By measuring usage frequency of individual products in successful evasions, the system can attribute effectiveness to specific scenarios independently, even when multiple alerts would normally lead to a single case, thereby preserving attribution information while maintaining detection coverage
Solution Approach 2:
The adversarial agent creates a copy or simulation of suspicious activity patterns to test the monitoring system. By analyzing which scenarios the agent bypasses in this simulated environment, the system can determine scenario effectiveness and attribution without the complexity of real-world multiple alert scenarios, preserving clear attribution information
3Measurement precision
If product usage frequency to evade scenarios is high, then loopholes in the monitoring system are exposed, but the system allows illicit activities to occur
Solution Approach 1:
The adversarial RL agent performs preliminary testing of the monitoring system by attempting evasions before actual illicit activities occur. By measuring product usage frequency in these preliminary evasion attempts, the system identifies loopholes and strengthens transaction constraints proactively, preventing illicit activities before they can occur in the live system
Solution Approach 2:
The system takes preliminary anti-action by automatically strengthening transaction constraints on products that the adversarial agent successfully uses to evade scenarios. This preemptive measure counteracts potential illicit activities by closing loopholes before they can be exploited, thereby reducing harmful factors while maintaining accurate loophole detection
Data Source
AI summary
Systems, methods, and other embodiments associated with automatic transaction constraint modification are described herein. In one embodiment, a method includes recording transactions by a reinforcement learning agent during attempts to evade one or more scenarios of a monitoring system. A usage frequency is determined for a product in a subset of the attempts to evade the one or more scenarios that are successful. The usage frequency is then compared to an expected usage frequency for the product. A transaction constraint on the product is then automatically modified to cause the usage frequency to be at or below the expected usage frequency.