Source-linked AI summary
A Classification Model for Sensing Human Trust in Machines Using EEG and GSR
Kumar Akash, Wan-Lin Hu, Neera Jain, Tahira Reid
TL;DR
Human–machine interaction requires trust estimation, yet a static model using psychophysiological signals remained a gap. This paper develops classifier-based trust sensors from EEG and GSR data, finding that customized feature sets improve mean accuracy over general features but require more training time.
Problem
A static model for estimating human trust from EEG and GSR remained underdeveloped, despite trust’s importance for human–machine interaction.
Method
Using data from 45 participants, the paper trains classifier-based trust sensors with EEG and GSR features, comparing general and participant-customized feature sets.
Results
Customized trust sensor models achieved 78.55% mean accuracy versus 71.22% for general models, with the comparison trading higher performance for longer training time.
Takeaways & Limitations
Real-time psychophysiological measurements can estimate human trust and support intelligent systems that respond to changes in trust behavior.
Takeaways & Limitations
The model includes only a subset of non-verbal signals, and individual differences such as dispositional trust factors may contribute to lower general-model accuracy.
Abstract
from arXiv · showhide
Today, intelligent machines \emph{interact and collaborate} with humans in a way that demands a greater level of trust between human and machine. A first step towards building intelligent machines that are capable of building and maintaining trust with humans is the design of a sensor that will enable machines to estimate human trust level in real-time. In this paper, two approaches for developing classifier-based empirical trust sensor models are presented that specifically use electroencephalography (EEG) and galvanic skin response (GSR) measurements. Human subject data collected from 45 participants is used for feature extraction, feature selection, classifier training, and model validation. The first approach considers a general set of psychophysiological features across all participants as the input variables and trains a classifier-based model for each participant, resulting in a trust sensor model based on the general feature set (i.e., a "general trust sensor model"). The second approach considers a customized feature set for each individual and trains a classifier-based model using that feature set, resulting in improved mean accuracy but at the expense of an increase in training time. This work represents the first use of real-time psychophysiological measurements for the development of a human trust sensor. Implications of the work, in the context of trust management algorithm design for intelligent machines, are also discussed.
1 INTRODUCTION
Intelligent machines increasingly collaborate with humans, making trust central to successful interaction. The paper addresses the need for real-time trust sensing by developing classifier-based models from psychophysiological measurements.
- Human supervision and intervention remain essential as intelligent systems become more common across sectors.
- Situational and learned trust can change during a single interaction, motivating feedback control that responds to changing trust levels.
- Continuous self-reports are impractical for feedback control, while psychophysiological signals had not been studied for real-time trust sensing.
- The paper presents a classifier-based human trust sensor using real-time GSR and EEG measurements collected in a human subject study.
- The proposed framework includes data collection, feature extraction, feature selection, model training, and model validation.
2 BACKGROUND AND RELATED WORK
Prior work linked EEG and GSR measurements with trust-related responses but did not establish a static, real-time trust estimation model using these signals. The paper focuses on noninvasive measurements suitable for real-time processing.
- EEG captures cortical brain activity whose changes reflect human thoughts, actions, and emotions.
- Earlier EEG trust studies focused on event-related potentials, which average repeated responses to specific events.
- GSR measures skin conductivity and reflects arousal through sympathetic nervous system activity.
- Although GSR values have been associated with trust and cognitive load, GSR-based trust estimation had not been explored as a real-time model.
- A remaining gap was a static model that estimates trust from EEG and GSR while remaining suitable for real-time implementation.
3 METHODS AND PROCEDURES
The study collected behavioral and psychophysiological data in a within-subjects human-subject experiment designed to identify trust-related features and build an empirical trust sensor model.
- The experiment used a within-subjects design to collect and analyze behavioral and psychophysiological data.
- The study sought psychophysiological features significantly correlated with human trust in intelligent systems.
- The empirical model-building process involved feature extraction and feature selection after data collection.
- Forty-eight adults aged 18–46 were recruited from West Lafayette, Indiana, and compensated at $15/hr.
EEG.
The study framework includes data collection and model-development stages, while the EEG-related procedure used GSR recordings from participants’ fingers as a psychophysiological measurement.
- The proposed framework proceeds through data collection, feature extraction, feature selection, model training, and model validation.
- Skin conductance was measured from the index and middle fingers of the non-dominant hand at 52 Hz.
GSR.
The experiment elicited changing trust by varying a simulated obstacle-detection sensor’s reliability, while collecting behavioral responses and psychophysiological measurements. Reliable and faulty trials were organized into counterbalanced phases to examine trust and distrust dynamics.
- Experimental setup: Participants wore EEG and GSR sensors while interacting with a simulated car obstacle-detection system.They evaluated sensor reports and selected trust or distrust before receiving correctness feedback.
- Trial procedure: Each trial presented either an “obstacle detected” or “clear road” report, followed by a trust/distrust response and correct/incorrect feedback.
- Reliability manipulation: Reliable trials used 100% accurate sensor reports, whereas faulty trials used 50% accuracy.Sensor faults were presented in randomized order.
- Experimental design: Each participant completed 100 trials divided into three experimental databases, with database ordering counterbalanced across two groups.The first two databases contained either reliable or faulty trials, and their lengths were selected to approach steady-state trust or distrust responses.
- Behavioral validation: Online validation responses showed high trust during reliable trials and low trust during faulty trials regardless of group.The hypothesis was validated using responses from 581 online participants, with 58 identified as outliers.
4 DATA ANALYSIS
The analysis pipeline cleaned EEG data, segmented signals into overlapping one-second epochs, extracted time- and frequency-domain features, and checked the expected trust pattern in online responses.
- Behavioral check: Online participants showed high trust in reliable trials and low trust in faulty trials regardless of group assignment.
- Preprocessing: EEG preprocessing removed artifacts and excluded participants with anomalous spectra, leaving 45 participants for analysis.The excluded artifacts included electromyography, electrooculography, spikes, saturation, and excursions.
- Signal segmentation: The study duration was divided into one-second epochs with 50% overlap to support real-time evaluation of rapid EEG changes.Only epochs associated with stimulus presentation were considered, and epoch counts varied with response time.
- Feature extraction: The feature set combined time-domain and frequency-domain EEG measurements because both domains are relevant to classifying cognitive tasks.
EEG.
EEG features were extracted from seven channels using time-domain statistics, inter-channel correlations, and wavelet-based frequency analysis. GSR analysis retained phasic activity to represent faster task-related responses.
- Time-domain features: Six time-domain features were extracted from seven EEG channels for each 256-sample one-second epoch.The channels were Fz, C3, Cz, C4, P3, POz, and P4.
- Channel interactions: The EEG time-domain set contained 42 channel-level features, with 21 additional features from pairwise correlations among seven channels.The correlations represented interactions between different brain regions.
- Frequency-domain features: Discrete Wavelet Transform decomposition was used to capture time-varying EEG frequency information with the Daubechies db5 mother wavelet.DWT provides better time resolution at high frequencies and better frequency resolution at low frequencies.
- Frequency-domain features: Mean, variance, and energy were computed across four decomposed bands and seven channels, producing 84 frequency-domain features.The total was 3 features × 4 bands × 7 channels.
- GSR features: GSR was decomposed into tonic and phasic components, but only the phasic component was used because the task operated faster than tonic changes.Maximum and net phasic components were calculated.
5 FEATURE SELECTION
The paper reduces a 149-feature psychophysiological set through filter and wrapper selection, then compares general and participant-customized feature-selection strategies for trust sensing. The customized approach reveals substantial individual feature diversity, while several EEG and GSR features recur across models.
- Selection algorithms: ReliefF initially shortlists features by local class-discrimination ability, and SFFS then selects the final subset using QDA misclassification rate.SFFS uses 5-fold cross-validation while iteratively adding features until performance no longer improves.
- Feature-selection rationale: Feature selection reduces the 149-feature set to avoid the curse of dimensionality in the limited-sample setting.The candidate features are evaluated as predictors of Trust or Distrust classes.
- Model variants: The two model variants select either a common feature set for a general population or a separate feature set for each individual.The distinction reflects differences in human trust behavior.
- Customized model: Customized selection produced an average of 4.33 features per participant, compared with 12 features when participant data were aggregated.The individualized feature sets showed substantial diversity across participants.
- Shared features: More than half of the most common customized features, including mean frequency at C4, were also significant for the general trust sensor model.This indicates overlap despite the diversity of participant-specific significant features.
- Significant features: Significant predictors included central-region EEG measures, left-parietal high-frequency EEG activity, and phasic GSR.The reported EEG patterns include C3, C4, and Cz correlations, beta-band activity, and mean-frequency measures.
6 MODEL TRAINING AND VALIDATION
The study validates quadratic discriminant trust classifiers using psychophysiological features and compares general and customized models. Customized models generally improve accuracy and sensitivity, while general models reduce training demands and support broader applicability.
- Model training: Quadratic discriminant classifiers predicted categorical Trust and Distrust classes from selected psychophysiological features.The classifier’s Trust posterior probability also provided a continuous trust-response output.
- Model validation: 5-fold cross-validation evaluated binary classification using approximately 129 samples per participant.Performance robustness was further examined across 10,000 random divisions of the data.
- Performance: 70.52±0.007% mean accuracy was obtained for training-sample participants, compared with 73.13±0.010% for validation-sample participants using the general model.The similar performance across participant groups supported applicability of the selected features beyond the training sample.
- Performance: 78.58±0.0005% mean accuracy was obtained with customized models, which used fewer than 5 features versus 12 for general models.For some participants, customized-model mean accuracy reached 100%.
- Performance: 82.56±0.007% mean specificity and 72.83±0.007% mean sensitivity were achieved by customized models across participants.General-model specificity ranged from 75.49±0.009% to 79.49±0.013%, while sensitivity ranged from 64.17±0.010% to 65.35±0.015%.
- Implementation trade-offs: The general model achieved 71.22% mean accuracy across participants, while customized models traded longer training for higher accuracy and smaller online feature sets.General models avoid feature selection during implementation, whereas customized models require an additional training-time selection step.
- Limitations: Dispositional trust factors and additional psychophysiological signals may improve accuracy because the current features represent only a subset of non-verbal trust-related signals.The study focused on situational and learned trust.
7 CONCLUSION
The paper concludes that real-time psychophysiological measurements can estimate human trust in intelligent systems. It compares general and customized classifier models, finding higher accuracy with customized features while identifying training time and future validation as important considerations.
- Conclusion: Psychophysiological measurements can estimate human trust in intelligent systems in real time.The intended use is to help intelligent systems respond to changes in human trust behavior.
- Conclusion: The customized trust sensor model achieved 78.55% mean accuracy using an individually selected feature set.The reported trade-off was higher performance alongside increased training time.
- Future work: Future work includes larger samples, dispositional-trust factors, simulated and immersive environments, and real-life testing.The paper also notes that feature-selection criteria could be adapted to different applications.