Behavioral Footprints: Digital & Physical Traces

• Behavioral footprints are observable traces of human and AI actions in various environments, enabling identification, prediction, and auditing.
• They are operationalized using vectors, tuples, and sequential policies across online networks, mobility data, and LLM provenance tracking.
• Methodologies extract and model footprints via feature engineering, similarity measures, and state-action policies, while addressing privacy and interpretability challenges.

Searching arXiv for relevant papers on behavioral footprints across privacy, social systems, mobility, and AI-agent behavior. Behavioral footprints are the observable traces left by humans or agents as they act within digital, physical, and sociotechnical environments. In online social networks, the term denotes traces produced by posting, commenting, content patterns, and interaction history; in privacy research it denotes app-usage or browsing fingerprints; in mobility and epidemiology it denotes the spatial imprints of movement; and in recent AI work it denotes procedural or alignment-induced signatures that distinguish models and agents by how they act rather than by what they output (Kowsari, 2022, Sekara et al., 2018, Zhou et al., 2022, Oderinwale, 15 Jun 2026). Across these literatures, behavioral footprints are used for identification, prediction, surveillance, personalization, planning, auditing, and provenance analysis.

1. Conceptual scope

The concept is not tied to a single modality. In Reddit-based user analysis, behavioral footprints are defined as traces users leave on online social networks via their activities, content patterns, and interaction history, including posting, commenting, links, time, and frequency (Kowsari, 2022). In cross-network profile matching, a user’s digital footprint is constructed from publicly available profile fields such as username, display name, description, location, profile image, and number of connections (Malhotra et al., 2013). In smartphone privacy work, the footprint is an “app-fingerprint,” a monthly binary vector indicating whether particular apps were used (Sekara et al., 2018). In web measurement, the fingerprint is an $n$-tuple of the most visited domains (Oliveira et al., 2023).

Other literatures use the term in domain-specific but structurally analogous ways. “COVID-19 footprints” are the recorded locations that COVID-19-infected individuals visited before diagnosis, aggregated spatially to characterize local transmission risk (Zhou et al., 2022). In environmental HCI, behavioral footprints denote the environmental impact of daily actions such as travel, food purchasing, and waste disposal, with situated visualizations proposed to make those impacts tangible in urban space (Jansen et al., 2024). In forensic biometrics, the term extends to physical footprint and gait evidence, where the human footprint is treated as a unique physiological trace and gait as a behavioral biometric (Nagwanshi, 2022).

A common feature across these usages is that the footprint is not the latent trait itself. It is the externally measurable residue of behavior, from which identity, propensity, state, risk, or provenance may be inferred.

2. Formal representations and metrics

Behavioral footprints are operationalized through vectors, tuples, similarity profiles, sequential policies, and spatial aggregates.

Domain	Representation	Key quantity
Cross-network profiles	Similarity vector	$$\langle \text{username}_{score}, \text{name}_{score}, \text{description}_{score}, \text{location}_{score}, \text{image}_{score}, \text{connections}_{score} \rangle$$
App and web usage	Binary vector or domain tuple	app-fingerprint; $m^n_u = (d^u_1, d^u_2, \cdots, d^u_n)$
Mobility and epidemiology	Weighted location aggregate	$V_j$ and station-level COVID-19 footprints
Sequential behavior	State-action policy	$\hat{\pi}(a \mid s)$
LLM provenance	Refusal-vector aggregate	$\hat{\mathbf{f}}$

In app-usage privacy analysis, the monthly fingerprint for user $u$ in month $t$ is

$$\mathbf{f}^{(u,t)} = [f_1^{(u,t)}, f_2^{(u,t)}, ..., f_A^{(u,t)}],$$

where $f_i^{(u,t)} = 1$ if app $$\langle \text{username}_{score}, \text{name}_{score}, \text{description}_{score}, \text{location}_{score}, \text{image}_{score}, \text{connections}_{score} \rangle$$0 is used and $$\langle \text{username}_{score}, \text{name}_{score}, \text{description}_{score}, \text{location}_{score}, \text{image}_{score}, \text{connections}_{score} \rangle$$1 otherwise. Re-identifiability is quantified with unicity, measured as the fraction of users uniquely identified by a subset of apps, and temporal drift is measured with Jaccard distance (Sekara et al., 2018).

In web browsing, the fingerprint is defined as

$$\langle \text{username}_{score}, \text{name}_{score}, \text{description}_{score}, \text{location}_{score}, \text{image}_{score}, \text{connections}_{score} \rangle$$2

with uniqueness computed as the proportion of users whose $$\langle \text{username}_{score}, \text{name}_{score}, \text{description}_{score}, \text{location}_{score}, \text{image}_{score}, \text{connections}_{score} \rangle$$3-tuple is unique, and re-identifiability evaluated across separate time slices (Oliveira et al., 2023).

In epidemiological mobility analysis, an infected person’s visit weight is distributed equally across the unique locations visited before diagnosis:

$$\langle \text{username}_{score}, \text{name}_{score}, \text{description}_{score}, \text{location}_{score}, \text{image}_{score}, \text{connections}_{score} \rangle$$4

and station-level COVID-19 footprints within radius $$\langle \text{username}_{score}, \text{name}_{score}, \text{description}_{score}, \text{location}_{score}, \text{image}_{score}, \text{connections}_{score} \rangle$$5 are

$$\langle \text{username}_{score}, \text{name}_{score}, \text{description}_{score}, \text{location}_{score}, \text{image}_{score}, \text{connections}_{score} \rangle$$6

This converts individual visit histories into a local exposure-density measure (Zhou et al., 2022).

In policy-based behavioral analysis, user activity is modeled as a sequential decision process. The empirical occupancy and policy are estimated as

$$\langle \text{username}_{score}, \text{name}_{score}, \text{description}_{score}, \text{location}_{score}, \text{image}_{score}, \text{connections}_{score} \rangle$$7

$$\langle \text{username}_{score}, \text{name}_{score}, \text{description}_{score}, \text{location}_{score}, \text{image}_{score}, \text{connections}_{score} \rangle$$8

Here the footprint is no longer a bag of events but a state-conditioned action distribution (Schneider et al., 2 Feb 2026).

In LLM provenance tracking, the behavioral fingerprint is an aggregate of layer-wise refusal vectors. Harmful and harmless prompt centroids are contrasted at each layer, normalized, averaged across selected layers, and L2-normalized to obtain a unit fingerprint vector $$\langle \text{username}_{score}, \text{name}_{score}, \text{description}_{score}, \text{location}_{score}, \text{image}_{score}, \text{connections}_{score} \rangle$$9, which is compared by cosine similarity (Xu et al., 10 Feb 2026).

3. Extraction and modeling methodologies

The extraction pipeline depends on modality but typically combines preprocessing, feature engineering, and supervised or self-supervised modeling. In Reddit footprint analysis, preprocessing removes punctuations and numbers, lowercases text, applies lemmatization, and removes stopwords. Features include number of posts, number of comments, number of links, number of YouTube links, number of image links, number of Twitter links, average number of words in posts, average number of words in comments, date of first post, and date of first comment. These non-personal attributes are mapped to personal classes such as joining date and Karma using SVM, Random Forest, and deep belief network classifiers, with parameter selection via Grid Search and Random Search and two mapping procedures, AP1 and AP2 (Kowsari, 2022).

In cross-network identity disambiguation, the methodology is explicitly feature-specific. UserID and display name are compared with Jaro-Winkler similarity; description uses TF-IDF cosine similarity, Jaccard similarity, and WordNet-based Wu-Palmer similarity; location uses substrings, Jaccard similarity, Jaro-Winkler, and geocoded distance; profile image uses Mean Square Error, Peak Signal-to-Noise Ratio, and Levenshtein Distance; and number of connections is compared through normalized difference or class index. These similarity vectors are then supplied to Naïve Bayes, k-Nearest Neighbors, Decision Tree, and SVM classifiers with 10-fold cross-validation (Malhotra et al., 2013).

Recent work emphasizes sequential and latent representations. Delivery-optimized discovery in behavioral user segmentation encodes page-view sequences with a Hierarchical Attention Network, assigns users to clusters with a “Selector,” maps behavioral segments to static targeting attributes with “Beh2Stat,” and optimizes reach under a budget using Slack Minimization, Barrier Method, or Augmented Lagrangian Method (Chopra et al., 2024). Search-log analysis goes beyond clicks by inferring “cognitive traces” through an Information Foraging Theory schema and a multi-agent Analyst–Critic–Judge workflow, with human experts providing calibration and adjudication (Zerhoudi et al., 27 Feb 2026). Encrypted network traffic is modeled with a transformer backbone, per-user adapters, and a sparse autoencoder to extract interpretable longitudinal features from hourly activity vectors (Mahmood et al., 2 May 2026).

Model-centered footprints introduce another family of methods. Behavioral fingerprinting of LLMs uses a Diagnostic Prompt Suite and an automated pipeline in which a stronger LLM acts as judge under prompt-specific rubrics (Pei et al., 2 Sep 2025). Coding-agent analysis compresses action sequences with Byte-Pair Encoding or PrefixSpan, selects vocabulary size by V-measure, measures behavioral distributions with entropy and Jensen-Shannon divergence, and exposes them through the ProcGrep library for trace queries and procedural reward specification (Oderinwale, 15 Jun 2026).

This suggests a methodological shift from surface-event counting toward compressed, structured, and interpretable representations of procedure, context, and deviation from baseline.

4. Empirical findings and application domains

In online identity resolution, behavioral footprints are strongly discriminative. Cross-network profile disambiguation between Twitter and LinkedIn achieved accuracy, precision, and recall of 98%, 99%, and 96%, respectively, with UserID and Display Name as the most discriminative features (Malhotra et al., 2013). In Reddit profile inference, mapping public behavioral attributes to joining date and Karma produced classification accuracy higher than 89%, with DBN reaching 0.89 accuracy for both Karma and joining date in the reported comparison table (Kowsari, 2022).

Privacy research shows that minimal footprints are often sufficient for re-identification. In smartphone app usage, four apps are enough to uniquely re-identify 91.2% of users using a popularity-based attack heuristic, and 99.7% have a unique fingerprint when considering the full 12-month period (Sekara et al., 2018). In web browsing, merely the four most visited web domains are enough to identify 95% of individuals, and 80% can be re-identified in separate time slices of data (Oliveira et al., 2023). These studies reject the assumption that large populations necessarily provide anonymity.

Urban and public-health applications use footprints as exposure or impact surfaces rather than identifiers. In the adapted node-place-mobility framework, stations with high node, place, and human mobility indices normally have more COVID-19 footprints in proximity, and a log-log OLS model attained adjusted $m^n_u = (d^u_1, d^u_2, \cdots, d^u_n)$0 over $m^n_u = (d^u_1, d^u_2, \cdots, d^u_n)$1 stations (Zhou et al., 2022). In activity-based carbon analysis, 10% of users account for 88% of visitation-based footprints and 71% of distance-traveled footprints, with high-income users having larger footprints due to lifestyle choices and low- to medium-income users’ footprints reflecting limited access (Rajput et al., 2023). In environmental visualization, a workshop in Bordeaux generated 67 situated visualization concepts organized around four purposes—facilitating eco-friendly behavior, informing purchasing decisions, increasing general environmental awareness, and increasing social awareness—and four mechanisms: data comparison, surfacing hidden information, making footprints concrete, and externalizing collective knowledge (Jansen et al., 2024).

Commercial, psychological, and health applications use footprints for segmentation and inference. Delivery Aware Discovery models improved delivery metrics, reduced budget spend, and maintained strong predictive performance in discovery under budget constraints (Chopra et al., 2024). Psychological profiling from transaction footprints yielded AUC values from 53% to 64%, with Random Forest outperforming linear models for most traits and facets and with global and local XAI used to interpret predictions (Ramon et al., 2021). Encrypted smartphone traffic, interpreted as a longitudinal behavioral signal, revealed that stress is primarily associated with stable between-person differences, loneliness with within-person variation, and sleep disturbance with a combination of both (Mahmood et al., 2 May 2026).

Physical and forensic applications remain distinct but related. A review of human footprint and gait reports recognition rates typically ranging from 85%–99%, a study of 24,000 footprints with a chance of two footprints matching of 1 in 1.27 billion, and the use of footprint and gait evidence for personal identification, forensic podiatry, and access control (Nagwanshi, 2022).

5. Privacy, interpretability, and governance

Behavioral footprints are persistently privacy-relevant because they remain identifying even when explicit identifiers are absent. App-fingerprint work states that “hiding in the crowd” is an illusion: uniqueness remains high at scale, fingerprints drift slowly, and seasonal variability further affects re-identification risk (Sekara et al., 2018). Browsing data shows that even limited observation—such as the Top 100 most visited domains—still yields 82% uniqueness with $m^n_u = (d^u_1, d^u_2, \cdots, d^u_n)$2 (Oliveira et al., 2023). Public profile aggregation across social networks demonstrates that deanonymization risk can arise from ordinary fields such as name, location, description, and image, not only from rare or covert signals (Malhotra et al., 2013).

Attempts to mitigate these risks reveal explicit trade-offs. Cloaking studies compare fine-grained cloaking of individual footprints with metafeature cloaking based on automatically generated higher-level categories. After digital traces are doubled, only about 21.5% of previously protected male users remain successfully cloaked under fine-grained cloaking, whereas 86.6% remain successfully cloaked under metafeature cloaking. The same work finds the expected trade-off between privacy and personalization: cloaking an undesired trait also partially conceals other desirable traits (Goethals et al., 2023).

Interpretability and governance mechanisms address a different part of the problem: how footprint-based inference should be audited. In personality prediction from spending data, global rule extraction is used to approximate model logic and local counterfactual explanations show that individuals are assigned to personality classes because of their unique financial behavior; the paper argues that XAI promotes accountability (Ramon et al., 2021). In LLM provenance tracking, locality-sensitive hashing and zero-knowledge proofs are proposed to transform a private behavioral fingerprint into a publicly verifiable, privacy-preserving artifact, making provenance claims auditable without revealing proprietary model weights (Xu et al., 10 Feb 2026).

A recurring misconception is that privacy risk is primarily a content problem. The literature here indicates that timing, rarity, sequence, aggregation, and latent alignment behavior can all be identifying even when semantic content is hidden or degraded.

6. Behavioral footprints in AI systems

Recent research generalizes the notion of behavioral footprints from human traces to machine procedure. In coding agents, trajectories are treated as programs, and agents are identifiable by procedural habits or “fingerprints”: a probe over procedural signatures attributes an unseen trajectory to the correct agent at 85.7% accuracy, far above the random baseline of 11.1%. Behavior is most similar between models from similar release periods and those that are distilled from one another, with a distilled student and its teacher having Jensen-Shannon divergence of 0.25 (Oderinwale, 15 Jun 2026).

In LLM evaluation, “Behavioral Fingerprinting” argues that benchmark scores do not capture a model’s intrinsic cognitive and interactive style. Using 21 diagnostic prompts across 18 models and an automated LLM-as-judge pipeline, the study reports convergence in core reasoning and divergence in alignment-related behaviors such as sycophancy and semantic robustness, along with default persona clustering around ISTJ and ESTJ (Pei et al., 2 Sep 2025). This directly contests the view that model scale or benchmark strength is sufficient to characterize deployment behavior.

White-box provenance work pushes the idea further. Refusal-vector fingerprints, derived from internal directional patterns associated with harmful versus harmless prompts, are reported to be robust to finetunes, merges, and quantization, and to achieve 100% accuracy in identifying the correct base model family across 76 offspring models (Xu et al., 10 Feb 2026). Platform-agnostic moderation research likewise shows that behavioral policies can outperform content models in detecting coordinated information operations: policy-based classifiers achieve a median macro-$m^n_u = (d^u_1, d^u_2, \cdots, d^u_n)$3 of 94.9%, compared to 91.2% for text embeddings, and permit earlier detection from short traces (Schneider et al., 2 Feb 2026).

Another extension is inferential rather than identificatory. Search-log analysis introduces cognitive traces—labels such as FollowingScent, ApproachingSource, DietEnrichment, PoorScent, LeavingPatch, and ForagingSuccess—and reports that adding these traces improves session outcome prediction on AOL from $m^n_u = (d^u_1, d^u_2, \cdots, d^u_n)$4 and $m^n_u = (d^u_1, d^u_2, \cdots, d^u_n)$5 to $m^n_u = (d^u_1, d^u_2, \cdots, d^u_n)$6 and $m^n_u = (d^u_1, d^u_2, \cdots, d^u_n)$7 (Zerhoudi et al., 27 Feb 2026). Synthetic-data work then inverts the direction of analysis: PersonaTrace starts from a structured persona, generates event forests, and produces realistic digital artifacts such as emails, messages, calendar entries, reminders, and wallet passes, with intrinsic evaluations reporting greater diversity and realism than existing baselines and better out-of-distribution performance after fine-tuning (Wang et al., 12 Mar 2026).

A plausible implication is that behavioral footprints are becoming a unifying abstraction for both observing agents and engineering them: they are no longer only residues to be mined, but also targets for simulation, auditing, control, and provenance.