Abstract: A counterfeit detection system provides an artificial intelligence (AI) platform that implements a Generative Adversarial Network (GAN) to classify an image as one of a fake or genuine item and integrates a Classification Activation Module (CAM) to refine counterfeit detection. The GAN may include a generator that generates simulated counterfeit images for a discriminator. The discriminator may be trained to identify faked items by learning from the simulated counterfeit images and/or images of actual faked items. The discriminator may implement a deep neural network of convolutional layers that each analyze a region of an image and produce a weighted output that contributes to the classification based on the analyzed region. The CAM may identify the regions and weights relied upon by the discriminator, provide corresponding heatmaps to subject matter experts, receive annotations from the subject matter experts, and use the annotations as feedback to refine the classifier of the discriminator.

Abstract: Implementations are directed to providing categorization of transactional data, and include actions of providing a plurality of word embeddings based on domain-relevant text data, clustering word embeddings of the plurality of word embeddings into a plurality of clusters, receiving, in real-time, transactional data representative of a transaction, providing a category that is to be assigned to the transaction based on the transactional data, and the plurality of clusters, processing the category, the transactional data, the text data, and the plurality of clusters using a semantic search to provide reason text data, the reason text data representing a reason for selection of the category assigned to the transaction, and storing the transaction data, the category, and the reason text data in a transaction database.

Abstract: A counterfeit detection system provides an artificial intelligence (AI) platform that implements a Generative Adversarial Network (GAN) to classify an image as one of a fake or genuine item and integrates a Classification Activation Module (CAM) to refine counterfeit detection. The GAN may include a generator that generates simulated counterfeit images for a discriminator. The discriminator may be trained to identify faked items by learning from the simulated counterfeit images and/or images of actual faked items. The discriminator may implement a deep neural network of convolutional layers that each analyze a region of an image and produce a weighted output that contributes to the classification based on the analyzed region. The CAM may identify the regions and weights relied upon by the discriminator, provide corresponding heatmaps to subject matter experts, receive annotations from the subject matter experts, and use the annotations as feedback to refine the classifier of the discriminator.

Abstract: This document describes systems, methods, devices, and other techniques for detecting procurement fraud in one or more procurement processes. In some implementations, a computing device receives input data representing one or more procurement processes, processes the received input data to generate a respective risk score for each procurement process, each risk score representing a likelihood that the respective procurement process is fraudulent, comprising processing the received input data using (i) one or more predetermined rules and scenarios, and (ii) atypical patterns data mined through unsupervised learning mechanisms, and provides, based on the generated risk score, output data indicating procurement processes that are likely to be fraudulent.

Abstract: Implementations are directed to providing categorization of transactional data, and include actions of providing a plurality of word embeddings based on domain-relevant text data, clustering word embeddings of the plurality of word embeddings into a plurality of clusters, receiving, in real-time, transactional data representative of a transaction, providing a category that is to be assigned to the transaction based on the transactional data, and the plurality of clusters, processing the category, the transactional data, the text data, and the plurality of clusters using a semantic search to provide reason text data, the reason text data representing a reason for selection of the category assigned to the transaction, and storing the transaction data, the category, and the reason text data in a transaction database.

Abstract: A self-learning system for categorizing log entries may be provided. A text classifier may identify a log description of a log entry in response to text of the log description being associated with indicators of a word model. A datafield classifier may generate a datafield metrics including an accuracy of the categorical identifiers representing the datafield. A metafield classifier may generate a context metrics for the context of the log entry, the context metrics including an accuracy categorical identifiers representing the metafields. A combination classifier may form a weighted classification set and select a categorical identifier as being representative of the datafield based on the weighted classification set. A categorical controller may identify new categories based on an analysis of the context metrics of the log entry.