{"type":"DefensivePublication","slug":"37f50e79","title":"A fraud detection system and method","abstract":"A fraud detection system and method are disclosed. The method involves receiving an item of content and generating a first semantic embedding for it. One or more second semantic embeddings corresponding to known fraudulent items are accessed. A semantic similarity score is calculated between the first embedding and the second embeddings. This score is compared to a predetermined threshold. If the semantic similarity score exceeds the predetermined threshold, the item of content is flagged as potentially fraudulent. This provides an effective approach to identify fraud based on semantic similarity.","body_markdown":"## Field\n\nField of Invention\n\nThe present invention relates to fraud detection, and more specifically, to systems and methods for detecting fraud using semantic embedding similarity.\n\n## Background\n\nBackground of the Invention\n\nFraudulent activities pose significant challenges across various industries, leading to substantial financial losses and erosion of trust. Traditional fraud detection methods often rely on rule-based systems or statistical models that may struggle to adapt to novel fraud patterns or complex data structures. Such systems can be limited in their ability to identify subtle or nuanced indicators of fraud, particularly when dealing with textual or semantic information. There remains a need for improved and more robust mechanisms to efficiently and accurately identify and flag potentially fraudulent items or activities.\n\n## Summary\n\nSummary of the Invention\n\nThe present invention provides a fraud detection system and method that utilizes semantic embedding similarity. In one embodiment, an item of content, such as a document, is processed to generate its semantic embedding. This embedding is then compared to semantic embeddings of known fraudulent items to determine a semantic similarity score. If this score exceeds a predetermined threshold, the item of content is flagged as potentially fraudulent, thereby providing an effective and adaptive approach to identifying fraud based on semantic content.\n\n## Detailed Description\n\nDetailed Description\n\nThe present invention relates to a system and method for detecting fraud utilizing semantic embedding similarity. In one embodiment, the system processes an item of content, such as a document or a 'paper' as described by the inventor, to assess its likelihood of being fraudulent.\n\nReferring to a generalized embodiment, a method for detecting fraud in an item of content comprises the following steps:\n\nFirst, an item of content is received. This item of content can be any form of digital data that can be represented semantically, such as a document, an article, a report, a transaction description, or any textual data for which fraud detection is desired.\n\nSecond, a first semantic embedding for the received item of content is generated. Semantic embeddings are numerical representations of textual data where the meaning of words and phrases is captured. These embeddings allow for mathematical operations to assess semantic relationships between different pieces of text. The generation of this embedding typically involves machine learning models trained on large text corpora, which can understand context and meaning.\n\nThird, one or more second semantic embeddings corresponding to one or more known fraudulent items are accessed. These known fraudulent items serve as a reference dataset of what is considered fraudulent. Their semantic embeddings are pre-computed and stored in a database or data store accessible by the system. These embeddings represent the semantic characteristics commonly associated with fraudulent content.\n\nFourth, a semantic similarity score is calculated between the first semantic embedding (of the item of content under review) and the one or more second semantic embeddings (of the known fraudulent items). This calculation determines how semantically close the item of content is to known fraudulent examples. Various similarity metrics can be employed, such as cosine similarity, which measures the cosine of the angle between two vectors in a multi-dimensional space.\n\nFifth, the calculated semantic similarity score is compared to a predetermined threshold. As described by the inventor, in one specific embodiment, this threshold is 0.85. This threshold acts as a decision boundary to classify the item of content.\n\nFinally, if the semantic similarity score exceeds the predetermined threshold, the item of content is flagged as potentially fraudulent. This flagging can trigger various downstream actions, such as sending an alert, initiating a manual review, or preventing further processing of the flagged item. This process allows for an automated and data-driven approach to fraud detection.\n\nIn a system embodiment, a fraud detection system comprises various components configured to perform the steps of the method. Such a system may include a receiver module for obtaining items of content, an embedding generation module for creating semantic embeddings, a data store for housing known fraudulent item embeddings, a similarity calculation module for determining semantic similarity scores, a comparator module for comparing scores against a threshold, and a flagging module for indicating potentially fraudulent items. These components can be implemented in software, hardware, or a combination thereof, operating on one or more computing devices.\n\n## Abstract\n\nA fraud detection system and method are disclosed. The method involves receiving an item of content and generating a first semantic embedding for it. One or more second semantic embeddings corresponding to known fraudulent items are accessed. A semantic similarity score is calculated between the first embedding and the second embeddings. This score is compared to a predetermined threshold. If the semantic similarity score exceeds the predetermined threshold, the item of content is flagged as potentially fraudulent. This provides an effective approach to identify fraud based on semantic similarity.\n\n## Claims (illustrative — NOT for monopoly)\n\n1. 1. A method for detecting fraud in an item of content, comprising:\n  receiving an item of content;\n  generating a first semantic embedding for the item of content;\n  accessing one or more second semantic embeddings corresponding to one or more known fraudulent items;\n  calculating a semantic similarity score between the first semantic embedding and the one or more second semantic embeddings;\n  comparing the semantic similarity score to a predetermined threshold; and\n  flagging the item of content as potentially fraudulent if the semantic similarity score exceeds the predetermined threshold.\n2. 2. The method of claim 1, wherein the predetermined threshold is 0.85.\n3. 3. The method of claim 1, wherein the item of content is a document.\n4. 4. A system for detecting fraud, comprising:\n  a receiver configured to receive an item of content;\n  an embedding generator configured to generate a first semantic embedding for the item of content;\n  a data store configured to store one or more second semantic embeddings corresponding to one or more known fraudulent items;\n  a similarity calculator configured to calculate a semantic similarity score between the first semantic embedding and the one or more second semantic embeddings;\n  a comparator configured to compare the semantic similarity score to a predetermined threshold; and\n  a flagger configured to flag the item of content as potentially fraudulent if the semantic similarity score exceeds the predetermined threshold.\n5. 5. The system of claim 4, wherein the predetermined threshold is 0.85.\n6. 6. The system of claim 4, wherein the item of content is a document.\n","published_at":"2026-06-16T09:30:11.851284Z","publication_url":"https://depatent.com/p/37f50e79","license":"CC0-1.0","authors":[],"anonymous":true,"cpc_codes":[],"ai_disclosure":"Drafted with AI assistance · not legal advice","proofs":[{"kind":"doi","state":"missing","value":null,"href":null},{"kind":"rfc3161","state":"missing","value":null,"href":null},{"kind":"archive","state":"present","value":"web.archive.org","href":"https://web.archive.org/web/20261030022615/https://depatent.com/p/37f50e79"}]}