Food fraud is a global problem that typically involves the dilution or mislabeling of food products, or ingredient substitution. In 2013, horse meat was found in many supermarket meals in Europe that claimed to contain beef, for example, while milk has often been found to be watered down in India to increase profits.
A 2021 study by the United Nations Food and Agriculture Organization cited a 2018 European Commission finding that estimated " the cost of food fraud for the global food industry is approximately EUR 30 billion" (about $30.5 billion) each year.
While chemical analyses can be carried out in a lab to authenticate food, traditional methods are often expensive, time-consuming, and require technical expertise. That is why researchers are aiming to develop new tools that harness artificial intelligence (AI) to enable rapid, inexpensive screening of food and beverages.
"It would be a very exciting scenario to have AI help us expand the reach and impact of chemical analyses," says Patrick Ruch, a research staff member at IBM Research in Zurich, Switzerland. "All of the intelligence can be on a smartphone or in the cloud."
Ruch and his colleagues have been working on a system to authenticate beverages called HyperTaste that uses a small, portable device called an electronic tongue (e-tongue), combined with machine learning. The e-tongue contains 16 sensors made of conductive polymers that can be thought of as taste buds; when dipped into a drink, the sensors pick up chemical information in the liquid that can be converted into a unique digital fingerprint measured as a time series of voltages. "We know that the signal that we're measuring is a unique indicator of what's inside the liquid because these polymers are interacting with all of the small molecules inside," says Ruch.
Machine learning then is used to make sense of the complex signal detected, for example to identify a specific brand of wine or its origin. In recent work, Ruch and his colleagues focused on wines and juices, training three different machine learning models to perform various recognition tasks using data collected with a sensor-equipped robotic device. The automated system dipped its sensors into nine different types of fruit juices several times, collecting 72 voltage time-series measurements. The process was repeated using 11 different types of Italian red wines to generate 110 measurements. "Nowadays, you can really quickly obtain the data needed for training with automation," says Ruch. "Within half a day to maximum a day, you have all the training data you need."
HyperTaste then was tested on a subset of the data that had been set apart for that purpose. When tested on nine different types of fruit juices, the system identified each with up to 97.3% accuracy. The system also could predict whether consumers would like the taste of juices stored for different lengths of time at high temperatures, with up to 93% accuracy (as assessed by a human tasting panel). HyperTaste was able to identify an individual wine with up to 99.1% accuracy, depending on the machine learning model used, and it could be classified as originating from one of four Italian regions with up to 98.2% accuracy. "Overall, the accuracy is very good, especially compared to the average taster," says Ruch.
Abigail Horn, a research computer scientist and assistant professor at the University of Southern California, thinks HyperTaste is promising for the detection of food fraud and adulteration, due to its high accuracy. Also, the technology's ability to predict consumer acceptability "may have implications for public health through real-time food safety assessment at point of consumption," she says.
HyperTaste's performance has improved considerably since it was first developed a few years ago by refining the training methodology and the representation of the digital fingerprint used for training. In 2019, an earlier version of the system was able to distinguish between four different types of mineral water with about 68% accuracy.
Ruch thinks the system could help labs that investigate food fraud accelerate their analyses.
The team is also developing the tool for other applications, such as analyzing acidity in ocean water, which has been tested on an autonomous ship that recently completed a trip across the Atlantic. "This type of technology can help the community more easily characterize substances and derive insights either for innovation purposes, to develop new substances, develop new recipes, or to verify the identity of substances," says Ruch.
Another team is aiming to tackle food fraud simply by using a mobile phone and machine learning.
In previous work, Hui Wang, a professor in the School of Electronics, Electrical Engineering, and Computer Science of Queen's University Belfast in Northern Ireland, U.K., and his colleagues developed a system using a near-infrared (NIR) spectrometer to collect data from food samples for authentication. However, the device itself costs over $18,000, which would be a significant barrier to wider use. They later switched to a less-expensive type of spectrometer, but that would still be an extra piece of equipment to carry around. "I asked myself if it was possible to use the existing hardware in your mobile phone to do the same thing," says Wang. "Almost everybody carries a mobile phone with them, so that would be quite useful."
Wang and his team have developed an app that can be used to analyze food or drinks using a smartphone. A food sample of interest is illuminated by a sequence of colors on a smartphone's screen, while its front-facing camera captures a video. Using computer vision techniques, the video is processed frame by frame to extract spectral information, which is used to train a machine learning model. The idea is that a specific food sample can be distinguished from others due to color differences. The particular sequence of colors used for illumination, which the team has patented, is key to the app's success. "Different sequences of colors give you different levels of performance, so we optimized that sequence of colors," says Wang.
In recent work, the team focused on the authentication of olive oil and milk. They captured videos of 160 samples of olive oil, some pure and some adulterated by different concentrations of vegetable oils, and trained three different machine learning models with the processed data. For milk, videos were recorded of 138 samples that varied in their fat content and could be classified as skim, semi-skim, or whole milk.
The models then were tested on olive oil and milk samples that it hadn't seen before. Pure and adulterated oils were distinguished with up to 96.2% accuracy, while all the milk test samples were correctly categorized into one of the three classes by two of the models. Wang and his colleagues were surprised at how accurate the results were. "Initially this was just a hobby and I thought designing an app would be quite interesting," says Wang. "But it turns out that this mobile phone app can be used to do some serious work."
The team has also tested the app on pure minced beef, and on samples that contained pork, with good results. The researchers now plan to train and test the system on spices, which are often prone to food fraud since they combine several ingredients.
Wang's goal is to help consumers tackle food fraud with his team's app. He plans to adapt the system to allow citizens to monitor their environment too, by enabling the testing of river water, or of city air for pollution levels. He currently is investigating the detection of respiratory viruses such as COVID-19 using a spectrometer, but he would like to create a phone app capable of the task at a later stage.
"If we can test for viruses using a mobile phone, that will be very significant," says Wang. "I would like to empower citizens who don't have access to sophisticated equipment."
Sandrine Ceurstemont is a freelance science writer based in London, U.K.
No entries found