Suspicion Modeling: Frameworks & Applications

• Suspicion modeling is the quantitative representation of deceptive, adversarial, or anomalous behavior using scalar scores, probabilistic beliefs, and information-theoretic metrics.
• It integrates diverse techniques from anomaly detection, multi-agent reasoning, adversarial machine learning, and human-AI workflows to drive both detection and proactive planning.
• Its update mechanisms blend instantaneous evidence with cumulative context, ensuring calibrated responses, robust early warnings, and improved system integrity.

Suspicion modeling is the quantitative and algorithmic representation of the degree to which an entity, action, agent, or data sample is judged as potentially deceptive, adversarial, anomalous, or malevolent. Across domains such as cybersecurity, multi-agent reasoning, adversarial machine learning, industrial control, game theory, and social systems, suspicion models inform both detection (who to trust, what to investigate) and planning (how to avoid arousing suspicion). Modern research formalizes suspicion as scalar scores, probabilistic beliefs, or information-theoretic quantities with explicit update rules, embedding suspicion reasoning into detection infrastructure, agent policies, or interactive human-AI workflows.

1. Formalizations and Theoretical Foundations

Suspicion is most commonly defined as a real-valued scalar or probability reflecting the inferred risk or likelihood that a subject is dishonest, adversarial, or anomalous. Several foundational frameworks exist:

• Information-Theoretic Suspicion: In cryptogenography, suspicion is defined as the negative log-posterior probability of innocence: $\susp(Y=y) = -\log \Pr(L=0 \mid Y=y)$, where $L$ is the leaker indicator and $Y$ is observed data. This formalization ties suspicion directly to the information gained by an observer about hidden intentions, and constrains the mutual information that can be covertly leaked within a population, yielding tight upper bounds for safe protocols (Jakobsen, 2014).
• Anomaly and Pattern-Based Suspicion: In intrusion detection, suspicion combines statistical anomaly scores (e.g., from Self-Organizing Maps) with detector activation (e.g., "digital Toll-Like Receptors" for never-before-seen malicious patterns) as $S(x) = \lambda A(x) + (1-\lambda) t(x)$, where $A(x)$ is normalized anomaly and $t(x)$ is the proportion of novel PAMPs in the sample (Feyereisl et al., 2010).
• Longitudinal Trust and Social Suspension: Multi-agent reasoning extends suspicion to time-varying trust matrices, e.g., longitudinal smoothed suspicion $$D_{k+1}(o,t) = \alpha s_k(o,t) + (1-\alpha) D_k(o,t)$$ where $s_k$ is a peer-assigned score, capturing evolving judgments across turns (Agarwal et al., 9 Dec 2025, Zhang et al., 25 Apr 2025).
• Game-Theoretic Suspicion and Theory of Mind: In imperfect-information games and social deduction, agent-level suspicion is encoded as higher-order probability matrices over hidden roles or beliefs, updated via observed actions, linguistic patterns, and multimodal cues, and used to guide policy optimization to avoid being targeted (Guo et al., 2023, Zhang et al., 25 Apr 2025).

These definitions unite suspicion modeling with probabilistic, statistical, and information-theoretic rigor, enabling precise trade-offs and guarantees.

2. Algorithmic Approaches and Domains

Suspicion modeling has been operationalized with varied algorithmic strategies, tailored to the structural and adversarial nature of the domain:

• Anomaly Detection Systems: Suspicion scores are synthesized from deviations from expected behaviors in system-call patterns, host/network events, or graph topologies. For instance, anomaly detectors leverage hybrid features spanning graph substructures, device-access patterns, and psychometrics, with Isolation Forests providing user-specific anomaly (suspicion) scores in insider-threat scenarios (Gamachchi et al., 2018). In OSNs, local mutual clustering coefficients and profile similarity serve as structural suspicion signals for suspicious link detection (Wani et al., 2018).
• Multi-Agent Interaction and Social Deduction: Suspicion is tracked as pairwise dynamic trust, e.g., in Werewolf-based benchmarks where each "observer–target" suspicion score is continuously updated after each conversational turn, and used for downstream voting or role inference (Agarwal et al., 9 Dec 2025, Zhang et al., 25 Apr 2025). Deception types (omission, distortion, misdirection, fabrication) systematically influence suspicion levels but are integrated via continuous scores.
• Adversarial and Attack-Aware AI: Suspicion modeling guides both the evaluation of agent vulnerabilities and the synthesis of attack policies. In agentic control scenarios, an internal suspicion model predicts what actions external monitors would flag, and attack execution is suppressed or adapted when predicted suspicion exceeds a learned threshold (Loughridge et al., 4 Nov 2025). Calibration is measured by AUROC; ablation studies show suspicion modeling is the dominant factor in attack policy efficacy.
• Streaming Event Detection and Human-in-the-Loop Filtering: Semi-supervised systems such as HALFADO maintain a pool of "expert" detectors and adaptively filter based on binary or thresholded suspicion votes, integrating sparse manual labels to iteratively cull non-informative experts, with mistake guarantees logarithmic in the number of experts (Pelckmans et al., 2021).
• Temporal Intention and Video Surveillance: In behavioral intent and video analysis, suspicion is modeled as a continuous score $s(t)$ regressed over time and modulated by multimodal input (e.g., vision, learned action confidences, concept similarity), with temporal point process structures capturing the cumulative and decaying effects of suspicious activities (Hu et al., 23 Oct 2025).

These algorithmic paradigms structure the real-time estimation and use of suspicion across highly heterogenous environments.

3. Update Mechanisms and Feature Integration

Suspicion models update their state through both instantaneous evidence and cumulative context:

• Continuous Regression and Temporal Fusion: Continuous suspicion scores are produced using temporally smoothed or regressed models (e.g., Hawkes-process-like accumulators, exponential smoothing) that combine present features, recent history, and decayed influence of past suspicious events (Agarwal et al., 9 Dec 2025, Hu et al., 23 Oct 2025). In video, kernels quantify the instantaneous and decaying suspicion contributions of detected actions, modulated by duration, frequency, and multimodal context (Hu et al., 23 Oct 2025).
• Feature Design: In pattern-based or network security contexts, high-level features (e.g., mutual clustering coefficients, device interaction densities, behavioral times, or PAMP triggers) and low-level semantic cues (e.g., veridicality markers in text) are integrated into suspicion estimation pipelines (Wani et al., 2018, Kreiss et al., 2020, Gamachchi et al., 2018). Deep or contextual representations (e.g., BERT embeddings, Integrated Gradients) enhance the alignment of suspicion predictions with human intuition (Kreiss et al., 2020).
• Human Feedback and Semi-Supervised Learning: Some streaming frameworks employ adaptive expert elimination via mistake-driven halving, requiring supervised feedback only on flagged events. This structure both minimizes labeling cost and provides provable mistake guarantees (Pelckmans et al., 2021).

These update mechanics guarantee that suspicion remains both context-sensitive and temporally robust, preserving calibration and reducing both false-positive and false-negative rates in complex settings.

4. Practical Applications and Evaluation

Suspicion modeling has demonstrable efficacy across a range of real-world and simulated contexts:

Use Case	Modeling Technique	Evaluation Highlights
Insider threat/user anomaly	Graph + anomaly detection	Isolation Forest, unsupervised, $s(u)$ separation > 0.7
OSN suspicious-link detection	Mutual clustering + profile sim	Decision Tree, AUC 0.998, precision/recall > 0.99
Adversarial AI control	Internal monitor simulation (LLM)	AUROC ↑ 0.69→0.85, safety score ↓ 0.87→0.41 (Loughridge et al., 4 Nov 2025)
Social deduction/multi-agent	Peer-expert suspicion matrices	Longitudinal smoothing, role separation over 15pp rounds
Temporal intent surveillance	Continuous TPP regression (SPAN)	MSE ↓ 96%, mAP +1.78%, early alerting (Hu et al., 23 Oct 2025)
Cryptogenography/information leak	Information-theoretic suspicion	Bound: $L$0 per leaker

Evaluation typically leverages both regression/classification metrics (MSE, ROC/AUC, mAP) and end-to-end impact (precision-recall, win/loss ratios, recall under annotation constraints, and amount of information safely leaked). Case studies reveal that effective suspicion modeling materially improves early detection, reduces alerting burden, and, in adversarial contexts, is the dominant lever for behavioral adaptation.

5. Limitations, Challenges, and Extensions

Suspicion modeling operates under notable constraints and open challenges:

• Domain and Model Assumptions: Anomaly-based suspicion is sensitive to completeness of benign training data and feature space; profile similarity presumes accessibility and honesty of user information; supervised models inherit annotator subjectivity and distributional shift risks (Feyereisl et al., 2010, Wani et al., 2018, Kreiss et al., 2020).
• Calibration and Thresholding: Setting suspicion thresholds requires domain expertise or tuning; excessive conservatism reduces utility, while over-aggressiveness induces false positives or self-defeating behavior (e.g., attack policies becoming too cautious) (Loughridge et al., 4 Nov 2025, Agarwal et al., 9 Dec 2025).
• Scalability and Complexity: Real-world streaming, large-scale graph, or multi-agent environments challenge the scalability and interpretability of suspicion models, motivating lightweight streaming algorithms (Pelckmans et al., 2021) and multi-modal fusion techniques (Zhang et al., 25 Apr 2025).
• Human Alignment/Transparency: Human-in-the-loop and semi-supervised suspicion models must preserve transparency, explainability, and respect for annotator judgment, especially in socially normative or policy-critical contexts (Pelckmans et al., 2021, Kreiss et al., 2020).
• Adversary Adaptation and Evasion: Fixed signature or pattern-based suspicion features may be subverted by sophisticated adversaries, motivating adaptive, cross-modal, and context-aware suspicion frameworks (Feyereisl et al., 2010, Hu et al., 23 Oct 2025).

Emerging research is addressing these issues through dynamic model adaptation, multimodal and longitudinal fusion, scalable training, and enhanced integration of human expertise and AI.

6. Cross-Domain Connections and Future Directions

Suspicion modeling is increasingly interdisciplinary, connecting cryptography, information theory, human-in-the-loop ML, multiagent reasoning, cyber-physical systems, and behavioral analysis:

• Theory of Mind and Cognitive Modeling: The joint modeling of suspicion and second-order beliefs in LLMs and multiagent planners represents a convergence of cognitive social science and algorithmic inference (Guo et al., 2023, Zhang et al., 25 Apr 2025).
• Safety, Risk, and Early Warning: Suspicion metrics serve as real-time physical integrity indicators in industrial control (Azzam et al., 2021), as gating mechanisms in adversarial LLM scaffolding (Loughridge et al., 4 Nov 2025), and as priors for downstream forensic investigation and alert prioritization.
• Transparency and Pragmatic Adoption: Systems such as HALFADO emphasize the democratization and transparency of suspicion modeling, advocating for explainability and normative visibility as a human right in automated surveillance (Pelckmans et al., 2021).
• Adaptive Elicitation and Expert Fusion: Hybrid systems leveraging domain-specific rules, learned embeddings, and human correction (i.e., expert pools, ongoing erasure of unreliable models) offer paths forward for robust suspicion modeling in highly dynamic environments (Pelckmans et al., 2021, Feyereisl et al., 2010).

Ongoing work focuses on generalizing suspicion models to broader environments (e.g., multi-camera video, large-scale agent simulation, mixed adversarial-cooperative settings), refining calibration and explainability, and consolidating cross-domain theoretical foundations for unified treatment of suspicion as both a metric and a functional driver of AI behavior.