Can TinyML Send Polygraphs to the Ash Heap of History? Our Fellow Sarah Falkovic Is Set to Find Out!
—Post written by Sarah Falkovic—
Can TinyML be used to design a small-scale lie detector model?
Lie detection has been a field of interest for about as long as the field of neuroscience itself, if not longer! However, lie detection in our current culture involves administration by a trained professional in the form of a polygraph. While not invasive, polygraph exams tend to take long and can be fallible, tending towards false positives when someone may be telling the truth. This, combined with differences in implementation between different specialists, make it a great example of a space for implementing machine learning.
So my goal is to assemble a lie detector using TinyML that can be implemented in a classroom setting for educational purposes. Here’s the prototype I built:
Historically, electrodermal response (sometimes called galvanic skin response) has been a commonly used indicator in lie detection. It’s exactly what I am starting with as the base modality for detecting lies. Electrodermal activity (EDA) measures the autonomic activity of sweat glands, which you can’t fake, as a sign of stress. The conscious load of lying induces an increased stress response, whereas truthful responses should show no increased stress in answering questions. Ideally, that is how you can tell apart truthful responses from deceptive ones. Assuming that the lie detector doesn’t lie too! But that’s why we’re putting TinyML into the equation!
I originally wanted this project to be based primarily in EDA for the model, as it is less invasive for kids compared to other modalities like EEG, where electrodes can be uncomfortable for some students. So I began by constructing a base EDA measurement tool below inspired by Electronics for Everyone on Instructables. Essentially, this device by itself allowed me to get a handle on the nuance of EDA while I waited for my proper Grove EDA monitor to come in. However, instead of machine learning, I implemented arbitrary thresholds that would activate a red light in the presence of lying. Here’s a clip of a humorous use of my original device:
While charming, my homemade model had some connectivity issues, and I was happy to move over fully to my EDA sensors. After that point, I began collecting data in a 5-card trial. Here’s how the trial goes:
- Before the trial begins, subjects secretly select one card from a series of 5 cards, without the examiner knowing.
- They are asked to respond “no” when asked if they had selected a card, irrespective of whether it was the secret card they chose.
Such a test is known as a concealed information test (CIT) as someone who was somehow ‘in the know’ about the secret information (i.e. the secret card) should have a different response when asked about that card compared to the control ones. While this brought forward some interesting data indicating an increased amount of stress when the subject is being asked questions they knew they are lying about, it led to many confounding factors. For example, we experienced dataset drift, making the reliability of what data I had – questionable.
There are other similar concerns with just using EDA- specifically, that of confounding factors like temperature. This is why I am also looking into other modalities to use alongside EDA. I am currently researching the use of P300 signals, which are an EEG response that flashes around 300 ms after being shown a recognized stimulus. Such a reaction is ideal for identifying those who may know information about a crime.
Paired with P300, however, is the issue of salience in crime simulation. Most lie simulators in research settings simply rely on using cards or small choices that aren’t as memorable as where polygraph and lie detection technology are normally used. Similarly, as P300 is an indication of familiarity, making events that induce lying memorable should increase the accuracy of the model I create. My current goal is to run some trials with mock thefts where subjects will be presented with pictures of objects, and hopefully will respond with recognition when shown the item they have stolen compared to the control items in a visual P300 exam.
I am a rising senior at Macalester College, where I am majoring in neuroscience. In my free time, I love to read science fiction, paint, and I lead my ethics debate club at my college. Aside from my hobbies, my background is both in neuroscience and STEM education. I assist in several neuroscience and biology courses at my college, primarily in helping first year students learn neuroscience and cell biology. However, the vast majority of my experience lies in youth STEM education. During the year, I normally help plan and teach STEM curriculum with the Science Museum of Minnesota for elementary students in several local low-income schools. My skill lies in working within constraints to make affordable and accessible educational tools.
The overall theme for this summer’s fellowship is working with TinyML, a small-scale machine learning technique that can be implemented on microcontrollers. We can apply on-device analysis to continuous low-powered data collection, which is incredibly useful for measuring common biometrics in neuroscience like electrodermal response (EDA) or electroencephalogram (EEG) data. This improves accessibility for both engineering and neuroscience education as well, as these microcontrollers are very low cost and don’t require expensive lab materials, making it affordable as an educational tool to buy in bulk.