Supply Chain Fraud
AS5553A and AS6081
“Supply Chain Fraud” is truly a nasty issue thriving in supply chain especially during this shortage season. Your purchasing colleague worked hard to survey his/her open market contact, eventually found some parts you need from an irregular channel. You knew there was risk and you were trapped in a dilemma: You had delivery commitment so the pressure piled, and you didn’t have enough resources to justify it in full scope. You struggled in the middle of associated failure cost and revenue simply because of this possible fraud. That played again and again and life should not be in this way.
Recently I read a blog on linkedIn for this topic by Forensic investigation and litigation support services at BDO USA LLC. It highlighted a “fraud triangle” framework which used to explain why the supply chain fraud happened.
The triangle consists of three sides: motive, opportunity, and rationalization.
So like scenario mentioned: You want to relieve from shipment pressure and now you have the chips so you tell yourself that we have no room to validate these ICs thorouthly, then nightmare comes true like product yields dropped dramastically. Any remedy we can take to save ourselves from such tragedy?
Surely there is no magic tricks but some industries best practices might point out the direction. For example, aerospace is such a safety critical industry and fraudulent parts could bring unexpected risk so SAE formed committee and released “SAE AS5553D-2022: Counterfeit Electrical, Electronic, And Electromechanical Parts(EEE); Avoidance, Detection, Mitigation and Disposition” to help organizations and apply through their supply chain by outlining the requirement like conterfeit EEE parts control plan, personnel training, parts availability, the purchasing processes, purchasing information, and the verification of purchased EEE parts. Some other requirements like investigations, material traceability and control, reporting and assessment are also included.
There is an importanct concept for counterfeit avoidance is the “Parts Availability”, several dimensions should be taken into consideration like parts lifecycle management, alternate and substitute parts so we don’t make ourselves in such dreadful situation. Basically the technical measures for detection would be focus on verification by information and testing. Additonal tests are always expensive. That’s the reason a lot of emphasis put on such information especially like traceability.
Hyperledger Fabric and Data Science
However records could be unavailable or suspected of being falsified, could we solve the suspected data in light of technologies today? Blockchain is the most well known technology considered as it’s nature of decentralized ledger and algorithm for authenticity, There are various projects sponsored by Hyperledger foundation which is the main contributor in blockchain ecosystem and Fabric is the project whcih focus on supply chain application since the core of the relationship performance is driven by transparency. However we should not be naive to believe that it’s ultimate solution to eliminate all fraud as security issue could live in different levels to give theft opportunities, risk-based thinking should be borne in mind.
So how we apply data science skill in risk identification which is the first sep of ISO 31010 risk assessment? There was a very interesting and inspiring example which successfully identified fraud for schools scores cheating in Chicago by data analysis. This case has been summarized in “Rotten Apples: An Investigation of the Prevalence and Predictors of Teacher Cheating” by Brian A. Jacob and Steven D. Levitt and published on the 2003 August edition of the Quarterly Journal of Economics.
The story happened in the last decade of last century in the States. To improve students achievement, a number of states or districts implemented programs to use student test scores to reward or punish schools. And the federal legislation, No Child Left Behind in 2001 enforced this practice and high incentives might seduce unscrupulous teachers and administrators in cheating behaviors like changing student responses on answer sheets, or providing correct answers to students. Recall the “fraud triangle”, we do see the story perfectly fits in this framework: There is motive for such illicit activities because of the incentives, teachers did have the opportunities to manipulate the test result and surely it’s easy to rationalize this behavior in a capitalism society. But the point is how to analyze the risk in an empirical and analytic way supported by data and fact?
The authors used extensive test data for grades 3 to 8 between 1993 to 2000 from Chicago public schools and analyzed these data on two types of indicators: unexpected test score fluctuations and unusual patterns of answers for students within classroom. Like in machine learning today, the approach started from domain knowledge to adopt reasonable feature sets and model needs to be not only trained but also validated.
Teacher cheating, especially in extreme cases, is likely to leave tell-tale signs.
So either from “time-series” or “anomaly detection” point of view, there would be clues summarized from extensive observation and come out suspicious pattern. For example, if cheating happened in certain year in one class, we might see the higher score like a peak then back to lower achievement since not manipulated by illicit behavior afterwards. Or “identical answer block” which violated the assumption that the individual incorrect answer should display in a more random way. They integrated four similar features to come out a score to represent such unusual patterns of answers for analysis. You can find the detail in the Appendix “The Construction of Suspicious String Measures”. In general, they could be described respectively below:
- The least likely block of identical answers given on consecutive questions in the classroom.
- The classroom average (across items) of the variance of student response for specific item within classroom across all test items.
- The variance (as opposed to the mean) in the degree of the correlation across questions within a classroom.
- The extent to which a student’s response pattern was different from other student’s with the same aggregrate scores that year.
With extensive classroom data, the relationship of these two factors was drawn in above figure. There is slightly positive correlation by the seventy-fifth percentile of suspicious answer strings so there might not be cheating as the assumption that the ratio of two measures kept constant if not cheating. However, you could see the dramastic increase on vertical axes after this threshold so highly likely cheating behavior existed in this popuation. Although there is no direct evidence to identify cheating but with such data analysis, in fact the suspicious cheating behavior was much prevalent than expected.
Again, it’s impossible to identify all fraud by any means. In light of data analysis with other skills like statistic distribution and probability and domain knowledge, “Rotten Apples” proved the powerful insight could be available once a successful model built. It’s also the inspiration origin of CloudIQC.
The mission of GLInB is to bring most value by virtualization of your supply chain quality function to fit for challenges in today’s business environment.
Originally published at http://glinb.com on July 23, 2022.
