Machine Learning is Here – Use It Wisely

Machine Learning (ML), a subset of Artificial Intelligence (AI), is already used effectively in such applications as medical diagnosis, image processing, classification, prediction, regression testing, and more. When considering the use of ML, there are generally two areas of risk: security of its use and compromising the validity of its findings.

Figure 1: The Industrial Internet of Things (IIoT) has many interconnections that can be more efficiently managed with machine learning incorporated into the system. (Image source: SlideShare.net)

AI security attacks are not new, but they continue to become more sophisticated. The attack surface, the number of points where an unauthorized user can covertly enter or extract data, can be exploited. Three areas have risk vulnerabilities: input data, algorithm design, and output decisions.

Machine learning is best achieved by accessing huge amounts of data from which it learns at a level that’s difficult for humans to achieve. Attacks include trained models poisoned by the creation of “backdoors” and the insertion of a malicious payload or trigger. In this segment of AI, algorithms are complex and unpredictable, are not subject to standards and regulations, and are based on proprietary data, making tampering even more difficult to detect.

There’s another area of risk besides security. Since machine learning models are created by humans, they’re subject to bias that can be built into the model. Data bias is dangerous and needs to be carefully managed. Managing bias is a very large aspect to managing machine learning risks.

Risk also involves insufficient data and the existence of good or appropriate data. The lack of variable data with sufficient data points to find the best inputs for optimal outputs can be a big problem. The data that makes up machine learning models should be varied across data types, timeframes, and other forms of variability.

Finally, output interpretations exist. There can be output misinterpretations. Models provide estimates and guidance, but it’s important to consider how a model was built, what assumptions were made, and what the output is telling you in order for interpretations to have value.

There are already many cases of what can go wrong, such as:

• Algorithms were blamed for the 6% fall of the British pound in two minutes during the 2016 Brexit referendum.
• Algorithms used by criminal justice systems across the United States predicting recidivism rates is racially biased.
• Many brain study results are questionable after erroneous statistical assumptions and bugs are found in functional magnetic resonance imaging (fMRI).
• When Bitcoin prices rapidly surged in 2017, hackers mined using Google cloud instances for free. Google Cloud’s anomaly detection system was in use for Google Cloud instances so that clients were warned of the compromise.

Machine learning systems are solving difficult problems. The fact that it can be impacted negatively on a security and accuracy level will be impacted by strides in the technology used and in the growth of the successful applications that use it.

Recent advances in machine learning

STMicroelectronics recently announced the first machine learning application on its STM32G4 by partner Cartesiam, a member of the Machine Learning ST Partner Program. ST launched STM32Cube.AI so that developers could easily train a neural network by collecting data before processing it in a neural network training framework on a PC to recognize specific activities, such as walking, running, or swimming. The output is then converted into a code that enables STM32 MCUs to recognize the activities.

Figure 2: Pictured is the STMicroelectronics SensorTile Evaluation Board. (Image source: STMicroelectronics)

Cartesiam’s NanoEdge AI runs the learning phase on the microcontroller. Engineers turn to this solution when they can’t create neatly pre-trained models for specific situations, but still desire to use machine learning to come up with smart solutions. The training phase is run on the MCU to learn the normal behavior of a device within its intended environment, then run inferences on the same MCU to detect and report behavioral anomalies.

With NanoEdge AI, developers easily integrate local AI training and analysis capabilities into C code, which is optimized for STM32 MCUs. In demonstrations, Cartesiam showed how its machine learning libraries could use STM’s SensorTile module, the SensorTile Evaluation Board (Figure 2), to learn the behavior of a BLDC motor through vibration analysis, then detect and report an anomaly thanks to the embedded STM32L4 MCU.

A machine learning core is also found on STMicroelectronics’ advanced sensors, such as the LSM6DSOX iNEMO. The core, combined with a Finite State Machine (FSM) and advanced digital functions, provides the ability to transition from an ultra-low power state to high-performance, high-accuracy AI capabilities for battery-operated IoT, gaming, wearables, and consumer electronics. Supporting typical OS requirements, it offers real, virtual, and batch sensors with 9 kbytes of RAM that are available for dynamic data batching.

While artificial intelligence and machine learning in all of its forms continues to amaze, it will be the combination of new applications, strength of output, and ability to maintain security that will continue to fuel its use.