FRAUD NOTIFICATION DETECTOR TECHNOLOGY ON NEURAL NETWORK MODEL

Abstract

This article is devoted to a comprehensive analysis of modern threats and vulnerabilities associated with text fraud in the digital environment. Both traditional rule-based signature methods and modern machine learning approaches, including zero-shot classification and hybrid architectures, are considered. The practices of social engineering, phishing, and financial fraud are analyzed to identify key patterns and markers of attackers. A lightweight system has been developed that combines the speed of rules and the semantic depth of neural network models, which has improved the detection efficiency. An important element was the risk assessment methodology, which took into account the content of messages and the context, which reduced the number of false positives to 3%, while maintaining high accuracy. The system architecture is implemented with cascade processing, which ensures optimal use of computing resources. Of particular importance is the mechanism of adaptive redistribution of weights between the components of the system, which allows you to maintain the quality of work with limited resources. A fallback classifier has been implemented to guarantee fault tolerance, maintaining accuracy in the range of 85-90%. The study highlights the need for a multi-layered approach that combines signature detection and semantic analysis. Prospects for the development of the system include expanding the rules, integrating with reputation systems, training to adapt to new fraudulent tactics, as well as creating a distributed infrastructure. Continued research in this area is important to combat evolving threats and create effective commercial solutions to protect against text fraud.