Interpersonal Deception Theory & Gaze Classifications

• Interpersonal Deception Theory is a framework that categorizes deceptive behavior by differentiating deliberate strategic control from involuntary leakage in nonverbal ocular cues.
• The ICE framework quantizes interpersonal gaze into a nine-region model using DBSCAN clustering and grid-based indexing for precise pattern detection.
• Empirical findings show that ICE-derived gaze metrics significantly enhance deception detection accuracy compared to conventional affective features.

Interpersonal Deception Theory (IDT) addresses the management of deceptive behavior in dyadic communication, emphasizing the interplay between deliberate and involuntary nonverbal cues. Within this context, the classification and quantification of ocular signals—especially eye gaze relative to a conversational partner—enable the empirical distinction between strategic modulation and behavioral leakage. The advent of algorithmic frameworks such as the Interpersonal-Calibrating Eye-gaze Encoder (ICE) has facilitated rigorous quantification of these cues, underpinning both deception detection and broader assessments of communicative competence (Tran et al., 2019).

1. Theoretical Grounding: Strategic Control and Leakage Behaviors

IDT asserts that deceivers actively manage both verbal and nonverbal channels to maximize credibility while minimizing suspicion. Ocular signals are dichotomized into:

• Strategic control: Deliberate behavioral modulation, e.g., maintaining prolonged eye contact or “staring to convince.”
• Leakage behaviors: Inadvertent cues resulting from cognitive load or emotional arousal, manifesting as gaze aversion (looking away or down, repeated fixation changes).

This framework predicts nonverbal signal variability contingent on the cognitive and affective processes underlying deception. A plausible implication is that quantification of gaze patterns provides a robust proxy for discerning intent and authenticity in social exchanges.

2. Quantization of Interpersonal Gaze Using ICE

The ICE framework operationalizes interpersonal gaze measurement by transforming raw camera-relative gaze angles into a nine-region discretization centered on the partner’s face. ICE utilizes a density-based clustering approach—specifically the DBSCAN algorithm—where the densest cluster identifies the region of visual engagement (RVE). The surrounding spatial field is tiled into a $3\times3$ grid, with region assignment per frame $i$ via grid cell indexing on the $(x_i, y_i)$ gaze coordinates.

Key operational steps include:

• Setting $\mathrm{minPts} = \lceil0.01N\rceil$ for $N$ frames.
• Iteratively adjusting $\varepsilon$ (cluster radius) until two clusters result with size ratio $\leq10$, or defaulting to $\varepsilon=0.001$.
• Extracting the bounding rectangle $RVE=[x_{\min},x_{\max}]\times[y_{\min},y_{\max}]$ and assigning each frame to region $r_i\in\{1,\dots,9\}$.

This quantized approach enables fine-grained detection of directional gaze shifts, distinguishing sustained on-partner contact (Region 5) from aversive behaviors including downward gaze (Region 8).

3. Behavioral Mapping and Statistical Classification

ICE-derived region frequencies $\mathbf{f} = (f_1,\dots,f_9)$, with $f_j = \frac{1}{N} \sum_{i=1}^N \mathbf{1}\{r_i=j\}$, serve as input for deception analysis. Region 5 frequency reflects sustained strategic gaze, while Region 8 quantifies gaze-down behavior indicative of leakage.

Statistical methods include:

• Independent two-sample t-tests on region frequencies with Bonferroni correction ($\alpha=0.05/9$).
• Effect size estimation via Cohen’s $d$.
• Logistic regression models (both $\ell_1$ and $\ell_2$ regularization), optimized by cross-entropy loss over 5-fold participant-unique cross-validation splits.

In comparative experiments, ICE region frequencies yielded higher classification accuracy ($\approx 0.64$) than affective features from Affdex or OpenFace-derived action units alone ($\approx 0.52$–$0.55$), with combined ICE + affective features slightly improved ($\approx 0.66$). The F1 metric for ICE validation, calibrated against IR trackers, reached $0.846$ (SD $=0.086$), demonstrating reliable mapping of interpersonal gaze signals.

4. Empirical Findings: Deception-Linked Gaze Patterns

Experimental interrogation setup (ADDR framework) involved video chat dyads with randomized “witness” and “interrogator” roles. ICE-based analysis revealed:

• Honest witnesses had significantly higher frequency of looking down ($\bar f_8^{\mathrm{truth}}=0.047$) compared to deceivers ($\bar f_8^{\mathrm{bluff}}=0.020$), with $p=0.0004<0.0056$ after correction and $d=0.79$, a large effect size.
• No other gaze region differentiated truthful from bluffing respondents after multiple comparison adjustment.

This pattern substantiates IDT’s prediction that cognitive stress from truthful recall produces leakage—manifested as downward gaze breaks—whereas bluffing is associated with increased strategic gaze maintenance toward the partner.

5. Methodological Extensions and Validation

ICE facilitates interpersonal gaze quantification without the need for specialized hardware or prior spatial calibration. Cross-modal validation shows:

• On-partner accuracy against IR tracker: $0.766$ (SD $=0.118$).
• ICE $f_5$ correlates with human-rated face-to-face eye contact ($r=0.37$ for individual ratings, $r=0.92$ for group means).

This suggests ICE reliably operationalizes interpersonal gaze features, matching ground-truth and human observer metrics, and enables practical deployment wherever conventional video is available.

6. Practical Implications and Application Domains

ICE refines IDT research methodology in several respects:

• Provides a directional, region-based discretization of interpersonal gaze beyond simple binary at-target labels.
• Enables real-time deception screening in video-mediated interrogations and e-interviews without dedicated gaze trackers.
• Supports automated coaching for social skills in clinical or organizational settings with robust, quantitative eye-contact feedback.

A plausible implication is the extension of ICE to multiparty scenarios, where multiple regions of visual engagement may be identified, and dynamic modeling of gaze transitions further enhances predictive analyses of deception or rapport.

7. Future Directions

Suggested research avenues include:

• ICE generalization to small group interactions, identifying and tracking multiple RVEs.
• Integration of temporal gaze dynamics for richer inference of communicative intent and deception.

The quantification and classification strategies outlined by ICE represent a significant methodological advance for empirical IDT studies, enabling large-scale, unobtrusive analysis of interpersonal deception phenomena under both laboratory and applied conditions (Tran et al., 2019).