Artificial intelligence based fraud prevention in digital advertising

Throughout 2022 and until the project end in the beginning of 2023 the project team focused on project sub-goal seven after successfully creating the AI-driven fraud detection of unknown fraud types until the end of Q3/2021. The project's final R&D sub-goal is to "develop AI-driven ad quality analysis" by developing algorithms that recognize and analyze ad assets and ad contexts. Firstly, we focused on understanding what type of ad asset we are dealing with e.g. image or video in order to go one level deeper and check the displayed ad asset for any deviations, discrepancies, and anomalies compared to the set of ad assets approved by the advertiser. Understanding if a publisher uses a non-approved ad asset to promote an advertiser's brand is essential to ensure brand protection and compliance. Once a non-approved ad asset is identified, our algorithms are designed to give the non-approved ad asset an ad asset score depending on the extent it deviates from any approved ad asset and classify whether or not it contains any restricted content categories such as pornography, gambling, and alcohol and drugs. Secondly, the project team conceptualized algorithms that analyze the ad context of an ad asset and check if it falls into a restricted content category, is likely to contain malware, or is generally of low quality. The individual ad asset score and ad context score are combined in a holistic ad quality score with fraud incident flagging capabilities based on predefined thresholds. Our analyzed data sets show that the use of approved ad assets in a non-approved ad context is the most common violation as it yields the highest return for the fraudsters while inflicting the greatest brand and financial damage on the advertisers. In the upcoming months until the market launch of the add-ons, we will continue refining the ad context recognition algorithms, especially when it comes to optical character recognition (OCR) to convert images of text to processable text and natural language processing (NLP) to understand patterns, characteristics, and sentiments in the text content.

The project's primary outcome is a powerful digital advertising fraud prevention solution. Over the next 24 months, it will be integrated into our end-to-end digital advertising software products, providing our advertisers with a convenient one-stop shop for automating efficient and profitable media buying. We anticipate that the benefits to our advertisers will be significant, leaving them with more money to invest and thus triggering positive social spillover effects such as more jobs, lower consumer prices, and better products and services. Furthermore, focusing on customer LTVs will drive down intrusive advertising and encourage better ads, making the internet a better, less invasive place. The project will benefit SMEs more as fraud disproportionately affects SMEs over large enterprises due to their lack of financial means to fight it. So, by making fraud prevention available to and affordable for companies of all sizes, we expect to stimulate growth and innovation, particularly among SMEs.

Fraud in digital advertising is a growing, worldwide epidemic that affects market players of all sizes and across all industries resulting in multi-billion dollar wasted ad spend, artificially inflated consumer prices, degrading digital user experience, and security breaches at unprecedented scale. This project aims to advance fraud prevention and customer lifetime value (LTV) prediction using artificial intelligence (AI) to prevent ad spending on malicious and low-quality traffic sources. The development of entirely new transparent, resilient, and adaptive fraud detection algorithms combined with accurate and timely predictive LTV modeling could redefine digital media buying and completely reshape the digital life of every company and every individual who is connected to the internet through any type of device. The most significant R&D challenges facing the project are largely related to engineering and artificial intelligence. The scalability required to train models using millions of data points at high speeds is difficult to design for. Another unique challenge is presented by the fact that millions of models will need to be stored for every client to achieve the most detailed level of insights possible. This type of scale is quite rare in machine learning and unprecedented in advertising technology. Also, developing algorithms that combine many artificial intelligence techniques to detect fraud based on fundamental traits and signatures so they are adaptable to all current and future manifestations requires a significant leap in several disciplines of machine learning. Both types of challenges will be overcome through the ingenuity of the project partners, each with a wealth of relevant experience and leading expertise in their respective core competencies.