Detecting Fake Suppliers using Deep Image Features

Abstract

The Observatory of Public Spending (ODP, in Portuguese) is a special unit of Brazil's Ministry of Transparency and Office of the Comptroller-General (CGU, in Portuguese) responsible for gathering managerial and audit information to support the work of its auditors. One of the most important tasks of this unit is to monitor government suppliers who have won procurement processes. Image analysis of the location of many of these suppliers revealed suspicious scenes, such as rural areas, isolated places or slums. These scenes could be an indicator of fake suppliers with poor capacity of delivering public goods. However, checking thousands of images in order to find suspicious suppliers would be very expensive. Our objective is to automatically distinguish images of valid supplier locations from arbitrary buildings and landscapes. We extract deep features from a collection of Google Street View images using a pretrained convolutional neural network (Places CNN) to classify supplier locations and show that these features can be well applied to the context of identifying valid suppliers, independent of the image perspective that was collected.