Companies are exposed to various risks both externally and internally. Advances in technology have improved methods for managing various risks, but fraud from within a company is even more difficult to detect and prevent than threats from outside, and remains a serious challenge for any company.
One of the effective methods for detecting internal fraud risks is email auditing. Digital communications such as email and, in recent years, chat, always leave behind some kind of evidence during the fraud process. If you focus on this and conduct email audits on a regular basis, you will not only be able to quickly find evidence when fraud is discovered, but also minimize risks by catching signs before fraud is discovered. These efforts protect companies and employees from potential incidents and lead to healthy business operations and development.
Here, we will explain the risks of fraud for companies, and introduce FRONTEO's "KIBIT Eye” algorithm will be explained.
Supervision
FRONTEO Inc.
Behavioral Information Science Institute
DEPUTY DIRECTOR GENERAL
Keisuke Tomiyasu
Obtained a Ph.D. (Science) through physics research.He is engaged in the fusion of condensed matter physics research and experiments and data analysis at universities and UK science and technology facilities.He researched and developed the algorithm of artificial intelligence KIBIT, and succeeded in major innovation.Currently, he is leading the improvement and creation of further algorithms, the design of KIBIT-equipped products, and business reforms.
Companies face a wide range of risks
Companies face a wide variety of risks, including business strategy risks, legal violation risks, security risks, and financial status risks. Examples include quality fraud, antitrust cartels, bribery, information leaks, accounting fraud, and violations of the Financial Instruments and Exchange Act.
On the other hand, interest in compliance has particularly increased recently, and strengthening compliance systems is essential from the perspective of corporate social responsibility. In the unlikely event that legal violations or fraudulent activities come to light, companies will lose social credibility and suffer losses such as decreased sales. Depending on the case, large fines and fines may be involved.
Corporate internal fraud countermeasures move from investigation in emergencies to audit (monitoring) during normal times.
In order for companies to maintain reliable management, it is necessary to have a system in place to detect and respond to signs of fraud as mentioned above at an early stage. In particular, there is growing awareness of the importance of routine audits, which detect signs before they occur.
Due to the diversification of work styles and digitalization, the importance of auditing (monitoring) as a preparation for normal times is increasing.
Normally, companies' countermeasures against fraud mainly involve "emergency" responses such as the discovery of fraud and litigation. In recent years, with the spread of remote work and the diversification of work styles due to the fluidity of employment, as well as the development and complexity of digitalization, the backgrounds and methods of fraud have become diverse. Therefore, auditing to detect and prevent signs of fraud even during normal times, or in other words, monitoring daily interactions and detecting signs of fraud at an early stage, has recently been attracting attention as an effective means of risk avoidance. Masu.
No fraud is a perfect crime, and the signs of fraud remain in communications such as emails and chats. In particular, e-mail is the central tool for communication both inside and outside the company, so it is the most important investigation target in fraud investigation cases, and has held a particularly important position among electronic evidence.
Email auditing AI that helps companies detect fraud risks
AI (artificial intelligence) has been attracting a lot of attention, including ChatGPT, which has become a hot topic in recent years, but the range of applications for AI is not limited to uses such as text generation. The greatest feature of AI is its ability to process large amounts of data, making it extremely useful in identifying fraud risks from vast amounts of email data.
A company's email server records a large amount of email logs, but fraud cannot be discovered just by storing them. In order to effectively conduct an email audit, you need a method and criteria for acquiring and analyzing logs to find fraud.
If AI is used to detect small emails that contain signs of fraud among a large amount of normal business communication emails, fraud detection during ordinary audits and searching for evidence emails in the event of a lawsuit will become overwhelmingly accurate. This increases the speed of response and, as a result, minimizes risks.
Advantages of using AI for email auditing to detect fraud risk
When AI is used for email audits, the judgment criteria are fixed, making it less likely that human-like judgments will vary or be overlooked, and through machine learning, it will be possible to automatically learn fraud patterns that humans cannot notice, improving detection accuracy. There are unique benefits. In addition to these operational benefits, the use of email auditing AI also has significant benefits that are essential for maintaining the sound management of a company.
Possible to respond quickly and prevent damage from spreading
The burden of audit work can be reduced because auditors only check emails with a high risk of fraud that are analyzed and detected by AI. Since audit work during normal times becomes more efficient, fraud risks can be discovered early and proactively dealt with before emergencies arise. In addition, by discovering warning signs through appropriate analysis, it is possible to prevent damage from spreading in the event of an emergency.
Protect people and protect organizations
By discovering fraud risks early, you can respond quickly, which means, for example, that you can protect people who are exposed to power harassment, and you can protect your organization by preventing the damage from spreading. Early detection of fraud risk goes beyond just detecting fraud to protecting people and organizations.
You can highlight your company's strengthened compliance
Utilizing AI as one of the measures to strengthen a company's compliance system will also serve as a check against fraudulent activities. This will also help to demonstrate to both the company and outside the company that we will not overlook any fraud by conducting thorough audits to prevent fraud.
Points to note and challenges when using AI for email auditing to detect fraud risks
General disadvantages of using AI to detect fraud risks include the burden of preparing training data (learning data) and the fact that the AI model becomes a black box.
The burden of creating training data is heavy.
The burden of creating and preparing training data for building AI models for fraud risk detection cannot be ignored. Building effective AI models requires a large amount of high-quality data sets, but the reality is that collecting and maintaining them is time-consuming.
AI judgment criteria may be opaque and difficult to interpret
The AI's decision process as to why the detected data was (or was not) determined to be fraudulent is unclear, and the results are sometimes difficult to interpret, leading to the risk that the reasons for the AI's decision may become a black box.
Deep learning, which has become popular in AI in recent years, is virtually impossible to check the processing in the middle layer (hidden layer), and it is difficult for the engineers who developed it to explain the processing and mechanisms that are occurring within the model. If the reasons for email audit judgments are unclear, and events occur that are beyond control even when attempts are made to correct erroneous judgments, trust in the system will decline and there is a risk that the original purpose of the audit may not be achieved.
“KIBIT Eye” realizes email auditing with natural language processing AI
Since the content of emails is text, or language, natural language processing AI is used for analysis. Natural Language Processing (NLP) is a technology that converts the natural language of unstructured data (not numbers) into "numbers" that can be handled by computers and then analyzes them.
The natural language processing AI installed in FRONTEO's email auditing tool "KIBIT Eye" uses a unique algorithm that is a type of machine learning and different from deep learning.
Evidence of the high accuracy of FRONTEO's email auditing AI tool "KIBIT Eye"
FRONTEO's email/chat auditing AI tool "KIBIT Eye" has many improvements to its algorithm, and as a result it achieves high accuracy.
In our own verification, when comparing the extraction rate (discovery accuracy) of data subject to audit when the viewing rate was 20%, the discovery accuracy of the predecessor system was 82.5%, but with KIBIT Eye, it was 99.5%, which is a significant improvement in analysis accuracy. Improvement was confirmed. This evolution allows us to better distinguish between the meanings of words, further improving the accuracy of fraud investigations.
In-house verification has also confirmed that KIBIT Eye's data detection ability is superior to Transformer-based generation AI such as ChatGPT and BERT. This is actually natural; while KIBIT Eye is optimized for discovering target documents using a small amount of training data, ChatGPT aims for natural sentences and is designed to "continue predicting the next word (token)." This is because it is optimized, and general-purpose AI does not exist (this can also be understood as a consequence of the no-free lunch theorem).
Algorithm that analyzes based on mathematical formulas, different from deep learning
The algorithm in "KIBIT Eye" is different from the deep learning that has become common in AI in recent years, and its main focus is on solving problems using mathematical formulas. "KIBIT Eye" does not have the "hidden layer" that is computationally expensive and becomes a black box in deep learning, and is basically composed of calculation formulas, so it is very easy to operate, and it is easy to understand how which elements affect the results. It also has the advantage of being highly explainable and tunable.
Why AI specialized in fraud risk detection can demonstrate high accuracy in audits
KIBIT Eye's algorithm is inspired by the color recognition of people with synesthesia, and the output results use hues to determine the degree of fraud. It allows you to intuitively and accurately capture minute information that would have been missed using conventional methods.
In terms of mechanism, we have further maximized the number of dimensions (features) to the necessary minimum and maximized accuracy from the AI email audit tool that we have provided so far. Other reasons for the high accuracy include improved word weighting, optimal handling of features, and mechanisms to suppress overfitting. Significantly improves detection accuracy and efficiency (patented*).
*Patent number: Patent No. 7376033
Reason 1: Appropriate weighting of words to identify fraud
One of the important aspects of natural language processing is the "weighting" of words, that is, how to express the importance of words in a document. When building a model from training data, KIBIT Eye's unique algorithm assigns positive weight to words related to fraud, and negative weight to words that similarly appear in documents not related to fraud. allocate optimally. For example, when a person looks at a document and determines whether it is related to fraud, if a person sees a particular word more often than others in a piece of paper, they judge that there is a possibility that there is a connection. What we call "weighting" is an attempt to express sensations in a mathematical formula.
These "weights" are ultimately aggregated and output as a score, so the method of assigning appropriate weights to words is an important factor that affects the accuracy of AI.
Reason 2: Maximize analysis accuracy by optimizing features even with a small amount of training data
"KIBIT Eye" can detect target documents with high accuracy even with a small amount of training data.
Select from the maximum number of feature candidates and analyze with the minimum number of features.
In an AI model, features are like each item of data when the characteristics of the data are expressed in a table.
"KIBIT Eye" analyzes all words (≒morphemes) in email documents as features. The performance of the model built from there is automatically verified, and accuracy is improved by removing features (morphemes) that do not significantly contribute to improving performance. Finally, a model is constructed using only the minimum number of features required.
In machine learning, the number of dimensions, that is, the number of feature items, is not necessarily better. For a learning model with high accuracy and generalization performance*, it is essential to select and analyze the optimal features for the purpose, that is, high-quality features.
*Generalization performance: Ability to achieve accuracy equivalent to teaching data even on unknown data
Use information other than text
Normally, in email auditing, the text in a document is analyzed using natural language processing, but with KIBIT Eye, information other than text data is also considered as a candidate for analysis. For example, information such as ``number of characters'' and ``number of lines'' in the email body falls under this category, but this is similar to how when a person reviews an email, their impression is influenced not only by the text of the email but also by its length. .
Another feature of AI is that there are extremely few hyperparameters that control the algorithm, which also contributes to the lightness of the AI's operation.
Reason 3: Model configuration that suppresses overfitting
If you try to improve the accuracy of AI, you will end up with "overfitting," which may be correct on training data but lacks versatility on actual operational data. However, in KIBIT Eye, overfitting is suppressed by using a linear-based learning model and turning ensemble learning on and off.
Learning model is linear based
When building a model with KIBIT Eye, that is, finding a function, select a function that approximates a linear model or generalized linear model, that is, a straight line graph.
As shown on the right side of the figure, if the model is a linear model, that is, if the graph is a straight line, it is expected that even with a small amount of training data, generalization performance will be high, overfitting will be less likely, and analysis accuracy will be improved.
Automatically adjust ensemble learning on/off
Ensemble learning is a method of training multiple models to achieve higher accuracy than a single model. However, there are cases where ensemble learning actually degrades performance, in which case ensemble learning can be automatically turned off and accuracy can be increased while suppressing overfitting.
Reason 4: Highlight high score areas
When investigating fraud using a general email audit system, a high score (degree of relevance to evidence) is given to "emails" that have a high degree of relevance to fraud. However, especially with long data, it was difficult to understand which parts of the email contributed to the high score.
KIBIT Eye not only looks at entire email documents, but also sentences and words, and identifies and highlights high-scoring areas within the document. In general, it is said that even if only the output of AI results is obtained, the reason for the decision is often unknown. However, the highlights can tell you why the AI gave the document a high score. For reviewers, they can intuitively know which parts of a document are particularly relevant, which improves explanation and makes review operations more efficient.
AI analyzes the entire document, sentences, and individual words to highlight specific areas. Being able to perform such high-speed extraction is technologically revolutionary.
Email auditing AI that combines the explainability required of future AI with past know-how
What is required of future AI is “responsible AI” with high explainability.
While the much-talked-about ChatGPT has the excellent feature of allowing exchanges in very human-like sentences, it is also known to cause inaccurate information to be output, known as hallucination. Nowadays, when there are unprecedented expectations for AI, there will be a need for the perspective of being able to explain the output results for the practical application of AI, that is, its social implementation.
"KIBIT Eye" is an approach that is designed with the minimum number of parameters with explainability in mind, and analyzes based on calculation formulas without using deep learning. Since you can trace which factors affected the output results, it is easier to find the cause of abnormalities and reduce maintenance costs. If AI is capable of explaining, its judgments will be reliable and the results will be easier for users to accept, and if the AI reaches an incorrect conclusion, it can be tuned, ensuring safety. In fact, in the operation of "KIBIT Eye," FRONTEO's customer success team works to optimize the model by tuning the causes of misjudgments, and assists companies in implementing AI.
"KIBIT Eye" combines FRONTEO's behavioral information theory and know-how with the latest algorithms.
Of course, high performance is required for AI, but at the same time, the value of AI is also increased if the underlying database is correct and based on know-how tailored to the purpose.
As a pioneer in fraud investigation, FRONTEO has been focusing on corporate litigation support and auditing using AI for over 10 years, and has supported companies in responding to fraud risks.
"KIBIT Eye" is an email auditing AI that is made up of FRONTEO's over 10 years of behavioral information research and auditing know-how, as well as in-house developed algorithms. By leveraging the knowledge we have accumulated through corporate audits we have supported, incorporating insights from behavioral information science, and combining this with a cognitive model algorithm that reproduces human sensibilities, we have achieved highly accurate judgments comparable to those of reviewers. Achieves fraud detection.
Email auditing is necessary for all companies that do not tolerate fraud.
As companies are required to further strengthen their compliance systems, they are required to take a stance that does not tolerate fraud. Unfortunately, internal fraud continues to occur in many companies because it is difficult to predict the risks before fraud occurs. Auditing emails and chats is a necessary measure for all companies.
While expectations are high for the use of AI in business, the reality is that there are generally high barriers to introducing AI.FRONTEO, however, aims to lower the hurdles to introducing AI by utilizing more accurate AI with less burden. We believe that this will lead to the expansion of social implementation of AI. In this respect, "KIBIT Eye" can be operated with a small amount of training data, so the barrier to introduction is low, and the AI learns and reproduces human judgment and tacit knowledge to quickly find the desired data from a huge amount of data. It provides optimal performance in situations such as.
