CYBERSECURITY AND FOOD DEFENSE

By Robert Norton, Ph.D., Professor of Veterinary Infectious Diseases and Coordinator, National Security and Defense Projects, Office of the Senior Vice President of Research and Economic Development, Auburn University; Marcus Sachs, P.E., Deputy Director for Research, McCrary Institute for Cyber and Critical Infrastructure Security, Auburn University; and Cris A. Young, D.V.M., M.P.H., Diplomate A.C.V.P.M., Professor of Practice, College of Veterinary Medicine, Auburn University and Adjunct Professor, College of Veterinary Medicine, Department of Pathology, University of Georgia

Beneficial AI: Safe, Secure, and Trustworthy Artificial Intelligence for Food Safety

AI can help ensure food safety by analyzing patterns and trends that might not be readily apparent to human analysts

Food, Ingredient, Cuisine, Dish

Image credit: MediaProduction/E+ via Getty Images

SCROLL DOWN

In a previous edition of Food Safety Magazine, the authors described in an article titled, "A Future View of AI-Enhanced Biosurveillance and Comprehensive Food Safety Programs,"1 a system of integrated sensors that could be used to further food safety and protect the bottom-line aspects of brand protection. One of the prominent features discussed was the fact that sensor arrays generate massive amounts and complexities of data.

To make data useful to the end user, it must be validated (meaning characterized by veracity), aggregated, and analyzed. Artificial Intelligence (AI) becomes essential because of the volume, variety, and velocity of the data. For AI to work properly, the algorithm developers need to work closely with the analysts using the systems to extract insight from databases and the subject matter experts that can validate the machine learning.

How AI Aids in Decision-Making and Analysis

Decision-making is often described as comprising four phases:

  1. Observe
  2. Orient
  3. Decide
  4. Act.

Developed first by Air Force Colonel John Boyd to describe how fighter pilots react to threats, the "OODA Loop"2 was widely reapplied to the business world, when it was realized that these processes can also be applied to other types of decision-making. The mental construct likewise can be used by decision-makers who desire to make better decisions more rapidly than the opposing party. In aerial dog fights and in business, being faster often translates to triumph when faced with challenge. AI helps decision-makers "tighten the loop," meaning truncating the time necessary to make right decisions. Over time, AI algorithms can also be refined, which moves the decision-maker closer to "near-real-time" or even "real-time" operational action.

AI algorithms have been developed for the purposes of "cueing and tipping." Cueing means that data (or observations) can be rapidly fed into the "orient" portion of the analytical cycle for further review by the analysts and subject matter experts (SMEs), and for the eventual development of alternative courses of action.

The orientation phase is designed to set the stage for the analysis efforts, making available facts that are then interpreted correctly with a set of preset principles (e.g., the presence of pathogenic bacteria is an unacceptable aberration in food safety). The orientation phase also acquaints operational authorities with existing situations (e.g., the reported temperatures on line 1 and 2 are correct) and the environment (e.g., all systems are operating in proper mode). Orientation also ensures the correct placement of the sensors (e.g., the optical sensor on line 1 is orientated perfectly perpendicular to the product moving down line 1). All data gathered in this phase is sent to a data set, which then can be used for further analysis and/or recordkeeping).

In parallel with this process, which in automated systems could occur in microseconds, tipping and cueing can occur between sensors. Suppose, for example, a lower-resolution sensor detects an anomaly somewhere in the system, but because of its resolution limits it was not able to make a definitive identification, or it requires a different orientation to make a definitive determination. The lower-resolution sensor (Sensor A) could then "talk" to a higher-resolution sensor (Sensor B) by generating an electronic "tip," thereby cueing or instructing the higher-resolution or differently oriented sensor to look at the object or event of interest.

Depending on the system and protocols, tipping and cueing can be done with or without the human in the middle. No human in the middle means sensor-to-sensor communications, such as occurs in automated systems. If, however, a human decision-maker is necessary for regulatory or business purposes, then the tip can be sent first to an analyst or SME for review and then passed on, cueing the second sensor to scan the object of interest. In such a system, the human would decide if a higher-resolution sensor should be "queried," meaning electronically instructed to gather the higher-resolution data.

The purpose of tipping and cueing sensors is to maximize granularity in the data while limiting its flow. Wide-open sensors create huge amounts of data, which, if uncontrolled, would swamp data storage with useless data. For instance, suppose Sensor A is looking at product entering a food processing system, while Sensors B, C, and D are looking at the products as they move through various processes. Lag times occur in any system, meaning the product movement is not continuous. It makes no sense to have sensors turned on when products are not flowing, since the data would be null, creating more "noise" from which the "signal" (the data needed) would need to be separated. AI can "learn" through the cooperative efforts of the engineers, analysts, and SMEs which operational targets (e.g., adulterants) are of significance and when to turn on and off the appropriate sensor.

The disadvantage to the "human in the middle" sensor action is that it is slower than a fully automated system. In these less-than-fully-automated systems, the process can incorporate "AI-enabled systems," making the speed of action faster than "human alone" systems (e.g., organoleptic inspection), but it cannot match a fully automated system where AI drives the tipping and cueing. It is important to note that all AI systems, whether assisted or fully automated, must be properly "trained" to correctly identify the target.

Triggers and alarms can also be incorporated in the second or "orient" phase of the process. For example, at some point within the food product processes, an aberration is detected. The system, again using AI to assist, can instantly respond with a trigger. The trigger could, for example, cause a temperature in a critical process to be slightly adjusted to stay within the prescribed settings or alarm, when the process is totally outside of specifications.

“AI can also assist in determining if economies of effort can be safely implemented… Imagine the financial savings if one or more food safety-related samplings could be modified or perhaps even eliminated along the farm-to-fork pathway of food, while decreasing the likelihood of foodborne illness.”
Tints and shades, Monochrome photography, Black, Black-and-white, Line, Style
Monochrome photography, Parallel, Black, Black-and-white, Line, White

Doing More with AI Data Analysis

Once through the orientation phase of the cycle, generated data can be examined by analysts and SMEs for use in the development of final products (reports, graphs, charts, etc.). This data can also be used for the production of courses of action, such as when the AI model is being trained. AI enables those in need of insight to receive it more rapidly and accurately, enabling faster decision-making and action.

Cueing and tipping is possible without AI, but it would significantly increase decision-making time. Analysis before AI could often take weeks, months, or sometimes even years for analytical conclusions to be fully developed and disseminated, whether by government using the traditional "intelligence cycle" or using similar approaches in business intelligence and business decision-making procedures.

The traditional "intelligence cycle" is most simply explained as a loop consisting of five phases:

  1. Planning and direction
  2. Collection
  3. Processing
  4. Analysis and production (reports, summaries, charts, etc.)
  5. Dissemination.

"Trained AI" truncates the first four phases, enabling more rapid dissemination to decision-makers. Consider "planning and direction" first. Suppose a human-only examination of your company's food safety-related data (perhaps using traditional statistical measures) resulted in a wrong conclusion. For example, the conclusion was "A," when in fact the better conclusion was "B." As a result of this incorrect conclusion, planning and direction were likewise wrong, or perhaps not as correct as they should have been. In other words, "kind of wrong" may mean that the resulting food products remain safe but do so at a higher cost than is achievable. AI can often "see" the totality of the data and analyze patterns and trends that might not be readily apparent to human analysts and statisticians. AI assists the correct decision-making in planning and direction. It also helps spot subtle anomalies along the way (think again of cueing and tipping). AI helps detect errors and may also be capable of developing new insights that might otherwise be missed.

AI can also greatly assist in data collection. Again, looking at the totality of the historical and currently generated data, analysts using AI may detect gaps or redundancies. For example, AI may determine the collection of data from a different part of the many processes that make up safe food production. AI can also assist in determining if economies of effort can be safely implemented.

In the authors' last article,1 we discussed an AI-driven sensor system that uses a farm-to-fork approach in food safety. Imagine the financial savings if one or more food safety-related samplings could be modified or perhaps even eliminated along the farm-to-fork pathway of food, while decreasing the likelihood of foodborne illness. Given that the food chain is a very complex "system of systems," it is highly unlikely that a SME alone could devise and test such a hypothesis, whereas well-trained AI might be able to identify acceptable modifications. Conversely, the AI might also conclude that alternative testing sites or strategies would provide more actionable data, perhaps further reducing the potential for contaminated products.

AI can assist greatly in data processing, reaching beyond even the most powerful, traditional statistical methods used with big data. Traditional big data statistical methods must be used in the preliminary model development and in training the AI, but those methods alone are no longer able to handle the volume and velocity of the data generated by a comprehensive food safety program, where for instance every generated data point is accompanied by geospatial and temporal data.

The authors' next column will examine how AI can be used in malevolent ways to harm food companies and the safety of the products they manufacture, and how problem solutions can be applied to mitigate this risk.

References

  1. Norton, R.A., M. Sachs, and C.A. Young. "A Future View of AI-Enhanced Biosurveillance and Comprehensive Food Safety Programs." Food Safety Magazine December 2023/January 2024. https://www.food-safety.com/articles/9110-a-future-view-of-ai-enhanced-biosurveillance-and-comprehensive-food-safety-programs.
  2. Mind Tools Content Team. "OODA Loops: Understanding the Decision Cycle." MindTools. https://www.mindtools.com/a3ldgz1/ooda-loops.

Robert A. Norton, Ph.D., is a Professor and Coordinator of National Security and Defense Projects in the Office of the Senior Vice President of Research and Economic Development at Auburn University. He specializes in national security matters and open-source intelligence, and coordinates research efforts related to food, agriculture, and veterinary defense.

Marcus H. Sachs, P.E., is the Deputy Director for Research at Auburn University's McCrary Institute for Cyber and Critical Infrastructure Security. He has deep experience in establishing and operating sharing and analysis centers including the Defense Department's Joint Task Force for Computer Network Defense, the SANS Institute's Internet Storm Center, the Communications ISAC, and the Electricity ISAC. 

Cris A. Young, D.V.M., M.P.H., Diplomate A.C.V.P.M., is a Professor of Practice at Auburn University's College of Veterinary Medicine and an Adjunct Professor at the College of Veterinary Medicine at the University of Georgia's Department of Pathology. He received his D.V.M. from Auburn University's College of Veterinary Medicine in 1994. He completed his M.P.H. at Western Kentucky University in 2005 and is a Diplomate of the American College of Veterinary Preventive Medicine.

FEBRUARY/MARCH 2024

Font, Line, Text