Hey everyone, this is Srinidhar Lakkaraju and I'm going to present AI powered dynamic risk scoring for e commerce transaction. This presentation introduces an innovative framework for dynamic risk assessment in e commerce fraud. detecting, using machine learning and advanced machine learning techniques. We will also explore how combining reinforcement learning, Bayesian network, and real time processing architecture addresses the modern fraud detection challenges and significantly improves the accuracy of detection while enhancing the customer experiences. As e commerce reaches unprecedented scale, close to 5. 7 trillion dollars in e commerce volume globally in 2024. The fraudsters have been deploying increasingly sophisticated attack strategies. The fraud attempts have been increased to 1. 8 percent of overall online revenue. The Redistone Static Square scoring systems have become dangerously inadequate, particularly In the face of a staggering 307 percent surge in account takeover attacks, while businesses invest heavily in fraud prevention, allocating close to 15 20 per customer, they face a critical challenge maintaining robust security. without compromising the seamless shopping experience that modern customers demand. And also we have seen close to 6. 4 billion losses in card not present fraud across the major online platforms. So our approach is the dynamic approach to the e commerce security. So we want to have four different phases of it. The first being the multi layer validation in which each transaction is passed through interconnected AI validation gates, which combines with the rule based checks along with the machine learning analysis. We use the advanced ML models where we are processing the 3, 700 data signals in real time, which includes transaction patterns. It's user behavior. and merchant risk profiles. We want to do real time behavior analysis where we want to see the advanced device fingerprinting and network analysis to detect suspicious patterns. VPN usage and the bot activities within milliseconds. And we want to continuously, evolve. So we want to have a self optimized models that automatically adapt to the new frauds tactics. while learning from the transaction outcomes and emerging threat patterns. Our cutting edge architecture delivers enterprise grade fraud prevention through three synchronized layers. Each component is engineered for maximum performance. seamlessly processing massive transaction volumes while maintaining strict latency requirements. This robust foundation enables real time risk assessment while automatically adapting new to new threats, setting up new standards for e commerce security. So these three layers are the data ingestion layer. In this layer, we process over 42, 000 transactions per second. with sub 200 milliseconds, ensuring a real time fraud detection at scale. The AI ML processing engine, which powers 1. 7 simultaneously user session through intelligent load balancing and automated resource optimization. And the third one is a dynamic threshold management. We execute 872 smart threshold adjustments hourly, which allows us to continuously adapt to the emerging. Fraud patterns. The reinforcement learning implementation is network aware. It has a network of strategies which is distributed system of 200, sorry, 2847 nodes which can process close to 47, 000 transactions per second. The sophisticated architecture has dramatically improved There is detection accuracy by 42 percent. The advanced feature engineering can go through a deep analysis of 3847 unique transaction characteristics or features which allow us to detect the fraud with 99. 4 percent accuracy Our system analyzes the monitoring pattern across 847 distinct feature to identify suspicious activity with unprecedented precision. The optimized risk reward function. Our holistic reward system balances multiple critical metrics including the false positive. The fraud detection, customer satisfaction, and network performance. This balanced approach has driven us a greater percent of improvement in our overall system effectiveness. The combination of vision network architecture. delivers industry leading fraud detection by combining massive scale real time processing with a sophisticated probabilistic model. The system not only processes transactions at unprecedented speed, but also continuously adapts to new fraud patterns, maintaining sub millisecond latency while achieving a very good So the real time processing is able to handle 47, 000 concurrent transactions per second. With being up, using the 10847 distributed nodes, we try to utilize the project interference, which kind of leverages these, interconnect decision nodes. managing the 12, 647 dynamic probabilistic relationships. We have the feature processing where we are able to analyze the 847 risk indicators in millisecond to achieve to a fraud detection. This allows us to have real time updates as well where we are adjusting all this 847k risk, indicators, and the probabilities. per second to have an understanding or create a base for emerging patterns. And then, the way we process these risk variables transactions using the advanced entropy reduction, we are able to identify the emerging patterns as well. The implementation of our AI powered dynamic risk score system has delivered exceptional improvements across all critical performance metrics. Most notable, we achieved a dramatic 33 percent reduction in false positive rates, dropping from 2. 4 percent to 1. 42%, while simultaneously increasing our fraud detection rate by 33 percent to reach industry leading 99 percent accuracy. Customer experience has been remarkable enhancements as well. The satisfaction score surging from 30 percent to 75 percent the improvement in the partial Attribute or this improvement is partially because of the processing time Where we have reduced our processing average time to 27 Close to now 600 milliseconds and Also have the reliability or the system validity to be close to four nines Ensuring that our service is active for our customers and it's not going down We have seen a challenge with our data imbalance since these fraudulence data are very or have an extreme balance where only 0. 12 to 0. 27 transactions are fraudulent. So making it much more difficult to inherently identify or detect any action. To solve that, we implemented a sophisticated broad line smooth algorithm combined with adaptive synthetic sampling to balance data representation. So we intelligently re sample the 3. 8 million transaction hourly to achieve a near perfect balance instead of having a data imbalanced where we have 51 percent legitimate cases and we have 48 percent fraudulent cases. While doing this allows us to increase our fraud detection accuracy. By 48 percent for the minority cases while Maintaining the exception low false positive rate 2. 3 percent. The real time processing architecture enables us to do high speed processing Able to have massive data handling provide real time updates and also increase the overall performance of the system we are able to now process close to 2, 000 concurrent transactions per second with industry leading Sub to 20 millisecond response time, ensuring the seamless customer experience, able to efficiently manage 4. 2 terabytes of transaction data per hour across distributed nodes while maintaining the enterprise grade 4. 9 system reliability. The advanced materialized view architecture. Refreshes every, 200 milliseconds or close to 200 milliseconds, delivering an 84% reduction in query latency for instant fraud detection. This allows us to have the real time updates and allows us to real time call, call the fraud data. We have able to have, receive, or cut the processing time by 67%. While delivering a boost in the transaction and throughput capacity. So that's a major improvement in our performance The model drift management. We continuously monitor all the features across these 847 model parameters Crossing 2. 8 million predictions per hour. We do a real time evaluation which allows us to achieve 99 accuracy in the drift detection with immediate response to the performance regression. We process these distributions across different sliding window anywhere from one hour to 90 days which allows us to have more defined path or more defined way to identify and define the accuracy and also detect the false positive or the fraudulent data. We also do A B testing, where we utilize the 12 parallel valuation groups, each of them processing 2. 7 million transactions daily, which allows us to come up with the newer detection and enable us to do the deployments in both A B fashion. In conclusion, our revolutionary AI powered risk scoring system has transformed the e commerce detection by processing 4. 2 terabytes of transaction data. hourly with 99 percent accuracy in drift detection. By combining reinforcement learning with Bayesian network, we have created a solution that not only detects frauds at a very unprecedented precision, but also scales dynamically to handle growing transaction volumes while maintaining a superior performance matrix. We're able to do an adaptive security, where we are able to advanced Risk assessment system in which you can automatically calibrate the security levels keeping the system up and also reducing the customer friction We have improved the accuracy. So we have achieved 47 increase in the fraud detection precision while maintaining a low 0. 3 percent false positive rate Which protects both the merchants and the customers We have enhanced the transaction processing time to sub 220 milliseconds response time and handled 1, 892 concurrent transactions with close to zero degradation in the system. And our system is future ready in the sense where we are able to self optimize the system monitors. monitor the 3847 features across these 87847 model parameters, ensuring us to continuously adapt to the new emerging fraud patterns, allowing us to keep in sync with the newer fraud patterns that have been emerging and able to take action accordingly. Thank you all. Hope you like the presentation.