WorksComputer Science
LABBA: LLM-Assisted Behavioral Biometric Authentication
Abstract
Behavioral biometrics have emerged as a promising complement to traditional physiological biometrics by leveraging dynamic human characteristics. These modalities offer advantages including continuous authentication, inherent liveness detection, and greater resistance to credential theft and other spoofing attacks. However, existing systems remain limited in adaptability, contextual awareness, and semantic reasoning as they primarily rely on similarity metrics, static thresholds, or supervised classifiers. As a result, they often struggle with real-world conditions such as behavioral variability, environmental changes, contextual changes, and adversarial manipulations. To address these limitations, we systematically integrate LLM-based contextual reasoning into behavioral biometric authentication. LLMs provide capabilities in contextual inference, sequential reasoning, multimodal understanding, and adaptive decision-making. Building on these strengths, we integrate a lightweight LLM-based reasoning module with a novel phone-blowing behavioral biometric technique to develop a context-aware authentication framework. To this end, we first collect blow-acoustic signals from participants and systematically extract spectral and temporal features. These features are aggregated and converted into comparable similarity representations, using either discrete rule-based categories or continuous normalized scores, which are transformed into structured prompts through prompt engineering. We experimentally evaluate the proposed framework using blow-acoustic data collected from 50 participants under diverse behavioral and contextual conditions. The proposed framework achieves 99.46% and 99.38% authentication accuracy in the rule-based and score-based modes, respectively, outperforming its non-LLM counterpart while maintaining strong robustness under dynamic real-world conditions.
Introduction
The rapid growth of digital services, mobile computing, and interconnected critical infrastructures has significantly increased the demand for secure, seamless, and user-friendly authentication mechanisms. Conventional authentication methods such as passwords, PINs, and security tokens suffer from several well-known limitations, including susceptibility to credential theft, phishing attacks, replay attacks, shoulder surfing, and poor usability due to credential memorization requirements [10, 4]. As a result, biometric authentication has emerged as a promising alternative for identity verification by leveraging intrinsic human characteristics that are difficult to replicate or share [20, 37, 18].
Traditional biometric systems primarily rely on physiological characteristics [3] such as fingerprints, facial appearance, iris patterns, and palmprints. While these modalities provide strong discriminative capability, they often suffer from several practical limitations, including vulnerability to spoofing attacks, lack of continuous authentication capability, sensor dependency, and privacy concerns associated with the permanent nature of physiological traits. In contrast, behavioral biometrics [14, 13, 21] has gained increasing attention as a complementary and, in some cases, alternative authentication paradigm that utilizes dynamic human behaviors such as typing dynamics, gait patterns, speech articulation, touchscreen interactions, mouse movement patterns, and airflow or breathing behaviors.
Despite substantial advances in behavioral biometrics, existing approaches remain fundamentally constrained by their reliance on static template or signal matching paradigms. Most such systems employ similarity-based verification models [36], fixed threshold mechanisms [6], statistical classifiers [23], or supervised machine learning [26] approaches that operate primarily at signal or feature levels. Although these techniques demonstrate promising performance under controlled experimental settings, their effectiveness often degrades significantly in real-world deployments characterized by natural behavioral variability [42, 33], dynamic contextual conditions [7, 39], and sophisticated adversarial threats [40, 34].
One of the primary challenges in behavioral biometric authentication is the highly dynamic nature of human behavior [42, 33]. Unlike physiological biometrics, behavioral patterns naturally evolve due to emotional state, cognitive workload, fatigue, stress, environmental conditions, device posture, interaction context, and long-term behavioral drift. For example, a user’s typing rhythm may vary depending on stress level or hand posture, while speech and breathing behaviors may change due to environmental noise, health condition, or physical activity. Similarly, airflow-based biometric interactions collected using smartphone microphones may exhibit variability caused by device orientation, ambient acoustic interference, user fatigue, or contextual interaction patterns. Such behavioral variability frequently causes conventional static threshold-based authentication systems to suffer from increased false rejection and false acceptance rates.
In addition to natural behavioral variability, most existing behavioral biometric systems do not incorporate contextual awareness into authentication decision-making [7, 39]. In practical deployments, authentication trustworthiness depends not only on biometric similarity scores but also on contextual factors such as device state, environmental conditions, user interaction history, temporal patterns, sensor reliability, and session-level behavioral consistency. However, most current biometric frameworks treat authentication as an isolated classification problem and fail to exploit contextual information that may significantly improve robustness and reliability. This limitation becomes particularly critical in continuous authentication and multimodal biometric systems where authentication conditions continuously evolve over time.
Another important limitation of behavioral biometric systems is their increasingly exposed to sophisticated adversarial threats [40, 34]. Attackers may attempt to imitate legitimate user behavior, inject synthetic behavioral traces, replay recorded interaction signals, or manipulate contextual conditions to bypass authentication systems. Existing anti-spoofing approaches primarily rely on low-level signal analysis [24, 41] or statistical anomaly detection techniques [16, 17], which often lack the ability to reason about long-term behavioral consistency, contextual coherence, or user intent. Consequently, conventional behavioral biometric systems frequently operate as static decision engines that make authentication decisions solely based on similarity scores without understanding the semantic and contextual meaning of observed user behavior.
Recent advances in Large Language Models (LLMs) [8, 43] provide a promising opportunity to address the static threshold-based decision making limitation of biometric systems. Because LLMs have demonstrated strong contextual inference, sequential reasoning, multimodal interpretation, semantic understanding, and adaptive decision-making capabilities across a wide range of domains. This has led to increasing interest in their integration with biometric systems [27, 2, 32, 11, 12, 35]. However, the application of LLMs in existing biometric systems is mainly limited to physiological biometrics, particularly in image-based modalities such as face, iris, and vein recognition [11, 12, 35]. Their application to behavioral biometrics, such as acoustic-based biometrics, remains largely unexplored.
In this work, we design an LLM-assisted behavioral biometric authentication framework that leverages the powerful reasoning capabilities of LLMs to enable context-aware and cognitively driven behavioral biometric authentication. Specifically, we first design a novel phone-blowing-based behavioral biometric technique that authenticates users based on their blow-acoustic signals. This modality captures rich behavioral and contextual characteristics, making it well suited for LLM-based contextual reasoning. During authentication, users interact with smartphone microphones through personalized airflow patterns and blowing behaviors, generating signals influenced by a wide range of behavioral and environmental factors, such as blowing style, airflow dynamics, user posture, ambient noise, microphone placement, and device orientation. Conventional static threshold-based methods often struggle to maintain reliability under such dynamic conditions, motivating the integration of LLM-based reasoning.
To address this challenge, we systematically extract spectral and temporal features from the collected blow-acoustic signals. The temporal features are further processed using bandpass filtering and Gaussian smoothing to suppress noise and unwanted frequency components. The extracted features are then aggregated and converted into comparable similarity representations, using either discrete rule-based categories or continuous normalized scores, and transformed into structured prompts through prompt engineering. These prompts are subsequently processed by the LLM-based contextual reasoning module to interpret behavioral patterns and generate context-aware authentication decisions.
As a result, the proposed framework dynamically adapts to diverse behavioral and environmental conditions, improving authentication robustness under real-world variability. By replacing rigid static-threshold mechanisms with LLM-driven contextual reasoning, the framework establishes a new paradigm for behavioral biometric authentication in which decisions are guided not only by similarity matching but also by intelligent reasoning over behavioral and contextual information.
To evaluate the effectiveness of the proposed framework, we collected blow-acoustic signals from 50 participants under diverse contextual conditions involving variations in environmental settings (e.g., in quiet rooms and noisy public environments), postural modes (e.g., in sitting and standing positions), and phone interaction dynamics with different blowing patterns. The collected signals are processed using advanced feature extraction techniques, such as Gammatone Frequency Cepstral Coefficients (GFCCs) [22], while the LLM reasoning engine analyzes contextual information and behavioral consistency to support adaptive authentication decisions. Experimental results demonstrate that the proposed LLM-assisted framework achieves an authentication accuracy of 99.46% and 99.38% under rule-based and score-based modes, respectively, both substantially outperforming the 99.09% achieved by the conventional static threshold-based approach. More importantly, the proposed LLM-based framework significantly improves authentication robustness against behavioral variability, dynamic environmental conditions and adversarial contexts compared to conventional static threshold approaches.
In general, by systematically integrating biometric similarity analysis with LLM-based contextual reasoning, the proposed approach transforms behavioral biometrics from a purely signal-level matching problem into a cognitive trust evaluation process capable of operating effectively under dynamic real-world conditions. Overall, this work makes the following main contributions:
We propose a novel LLM-assisted behavioral biometric authentication framework that systematically integrates LLM-based contextual reasoning with behavioral biometric techniques to enable adaptive authentication under dynamic behavioral and contextual variations.
We evaluate the proposed framework using a novel phone-blowing-based behavioral biometric technique that exhibits diverse behavioral and contextual variations.
Experimental results on a dataset collected from 50 participants demonstrate that the proposed approach achieves 99.46% and 99.38% authentication accuracy under rule-based and score-based modes, respectively, outperforming its non-LLM counterpart while significantly improving authentication robustness under dynamic real-world conditions.
Threat Model and Security Goals
Threat Model
In behavioral biometric (or general) authentication systems, the primary objective of the adversary is to impersonate the legitimate user and gain unauthorized access to the protected system by bypassing the authentication system. Therefore, in our threat model, we consider different attacks that commonly target behavioral biometric systems, which includes:
Threshold-based attacks: Generates near-threshold behavioral patterns to bypass the authentication.
Mimicry attacks: Imitates user’s behavioral patterns.
Template compromise attacks: Compromises stored biometric templates.
Replay attacks: Replays previously captured behavioral signals.
Synthetic attacks: Generates artificial data using AI models and inject it to the authentication system.
Model-based attacks: Targets trained LLM models through techniques such as prompt injection and jailbreaking.
Security Goals
The proposed LLM-assisted behavioral biometric authentication framework focuses on improving authentication robustness under varying environmental, behavioral, and contextual conditions. We aim to address the limitations of conventional static threshold-based behavioral biometric techniques that often fail under dynamic real-world conditions due to natural behavioral variability and contextual changes. Therefore, the main objectives of this work are to achieve the following aspects.
Accuracy: As a behavioral biometric technique, we aim to achieve high accuracy to effectively minimize the risk of impersonation attacks (i.e., by minimizing false positives), while also enhancing user convenience by lowering the number of required authentication attempts (i.e., by minimizing false negatives). This requires accurately distinguishing legitimate users even under dynamic contextual conditions.
Resilience against known attacks: We aim to achieve high resilience against the common behavioral biometric attacks mentioned above.
Usability: Usability is a primary challenge in behavioral biometrics. In this work, we aim to achieve the key usability issues, including:
User convenience: We aim to achieve system’s intuitiveness (easy to use) and nonintrusiveness (not causing discomfort).
Contextual robustness: We aim to achieve effectiveness of the system under different real-world conditions and contexts, including in different postural modes (sitting and standing), environmental conditions (quiet and noisy), blowing intensity (low and high), and blowing patterns (simple and complex).
LABBA: Proposed Work
This section presents the proposed LLM-assisted behavioral biometric authentication (LABBA) framework designed to enhance authentication robustness under dynamic contextual conditions. The framework integrates a blow-acoustic based behavioral biometric feature extraction process with LLM-based contextual reasoning to enable adaptive and context-aware authentication decision-making. In brief, the user first blows toward the phone screen (as illustrated in Figure 1) where blow-acoustic signals are captured via the phone’s microphone. The captured signals consist of rich spectral and temporal features. Spectral features are extracted in the form of Gammatone Frequency Cepstral Coefficients (GFCCs) [22] and subsequently passed to a trained Convolutional Neural Network (CNN) model to generate discriminative embedding representations. Concurrently, temporal features are extracted using a segmentation-based algorithm to capture user-specific blow interaction patterns. The resulting spectral and temporal comparisons between the enrolled and query samples are then combined into a unified biometric representation, which is provided to a lightweight LLM for contextual evaluation and authentication reasoning. Unlike traditional behavioral biometric systems that rely on static thresholds and fixed similarity-based verification, the proposed approach dynamically interprets behavioral biometric patterns alongside contextual information to improve resilience against behavioral variability and adversarial manipulation. Here, contextual information refers to the user’s enrolled interaction history, the temporal blow patterns of the query, and its session-level consistency with the user’s own baseline, expressed through the spectral and temporal comparisons between the query and enrolled samples. The LLM reasons over this contextual information and returns an accept or reject decision together with a justification, while environmental conditions are reflected only through these comparisons and are not provided as separate inputs.

Figure 1. Phone blowing illustration
Workflow
Figure 2 illustrates the overall workflow of LABBA. The framework operates through a multi-stage pipeline that integrates data collection, feature extraction, feature aggregation, and LLM-assisted evaluation to perform biometric authentication with contextual reasoning capabilities. Initially, raw acoustic signals are collected by the phone’s microphone as users blow toward the phone screen, as illustrated in Figure 1. The acquired recordings are subsequently

Figure 2. A high-level workflow of LABBA - Similarity Computation uses either Rule-Based or Score-Based, not both. Red dotted arrows indicate the enrolled-data comparison paths.
processed using two complementary feature extraction mechanisms, namely spectral feature extraction and temporal feature extraction. Spectral feature extraction captures frequency-domain characteristics of the blow-acoustic signals, while temporal feature extraction models the structural and dynamic behavioral patterns present over time. The features of each attempt are then compared against the user’s enrolled data, which consists of blow samples recorded from the legitimate user during enrollment. The resulting spectral and temporal comparisons are kept together for subsequent processing. Following feature aggregation, a prompt generation module constructs structured prompts tailored to the target LLM model by incorporating the unified feature representation. The generated prompts are subsequently utilized for LLM-assisted authentication and reasoning-based evaluation.
Blow-Acoustics Data Acquisition
The data acquisition stage focuses on capturing airflow-induced acoustic signals generated during phone-blowing interactions. In the proposed system, users perform blowing gestures toward the smartphone microphone screen, producing acoustic patterns that contain user-specific temporal and spectral characteristics suitable for behavioral biometric authentication. The phone-blowing interaction is performed within a 5-second duration, during which users may exhibit personalized blowing behaviors, including variations in the number of blows, duration of each blow, blowing intensity, blowing patterns, and airflow dynamics.
Blow-acoustic biometric signals provide several advantages, including strong resistance against spoofing attacks, implicit liveness detection, continuous and unobtrusive authentication capability, non-invasiveness, high usability, and low-cost deployment through existing smartphone microphones. However, as a soft biometric modality, blow-acoustic signals are highly sensitive to contextual variability, including ambient acoustic noise, smartphone orientation, user posture (e.g., standing and sitting), variations in blowing intensity, environmental dynamics, as well as user fatigue and behavioral drift. These challenges motivate the integration of LLM-based contextual understanding and logical reasoning into the authentication process to improve robustness under diverse real-world conditions.
Feature Extraction
The proposed framework extracts discriminative behavioral features from blow-acoustic signals to characterize user-specific airflow interaction patterns. The extracted features include two main categories: spectral and temporal.
Spectral Feature Extraction Spectral features are extracted from the blow-acoustic recordings using a GFCC-based processing pipeline to capture discriminative frequency characteristics and short-term temporal variations. We adopt the GFCC configuration of BlowLive [9], the audio signal is segmented into overlapping frames of with a hop size. For each frame, a 24-channel gammatone filterbank is applied to obtain auditory-inspired spectral energy representations. The filterbank energies are then logarithmically compressed and transformed using a discrete cosine transform (DCT), where the first 13 coefficients are retained as the static GFCC representation:
where denotes the n-th frame (), is the vector of 24 gammatone filterbank channel energies at frame , and is the resulting static GFCC vector.
To further capture local shifts, first-order delta coefficients are computed using a symmetric regression window and concatenated with the static coefficients, resulting in a 26-dimensional feature vector for each frame.
where denotes the 26-dimensional feature vector at frame and denotes the corresponding first-order temporal derivative (delta coefficient).
To obtain stable and discriminative embeddings that are independent of temporal sequence length, the extracted GFCC sequences are subsequently processed using a CNN-based feature encoder. The encoder was trained on 70% of the sessions and evaluated on the remaining 30%, so that test sessions are unseen during training. It was optimized with a triplet loss, which pulls embeddings of the same user’s sessions together and pushes those of different users apart. The model consists of stacked 1D convolutional layers with batch normalization and Rectified Linear Unit (ReLU) [29] activation, followed by temporal mean pooling and fully connected layers. The final output is L2-normalized to produce a compact 128-dimensional spectral embedding representation for each blow-acoustic sample.
where denotes the sequence of frame-level GFCC feature vectors over temporal frames, and represents the CNN-based feature encoder that maps the variable-length input sequence into a fixed-dimensional embedding space.
The extracted spectral embeddings are compared against enrolled user templates using cosine similarity. Suppose and represent the embeddings (Eq. (3)) extracted from the enrolled and query blow-acoustic signals, respectively. The cosine similarity score between the two is computed as:
where is the enrolled blow-acoustic embedding, is the query blow-acoustic embedding, is the embedding dimension, and denotes the resulting cosine similarity for the current authentication attempt.
In the conventional approach, this similarity score alone is used to make authentication decision. However, this score alone is insufficient for robust authentication under dynamic contexts and conditions discussed above, which motivates our use of LLM-based contextual reasoning using temporal features, as discussed below.
Temporal Feature Extraction Temporal and blow-dynamic features are extracted to characterize user-specific blowing patterns. Given a 5-second blow-acoustic recording, the signal first undergoes bandpass filtering to suppress unwanted frequency components. The filtered signal is then full-wave rectified to obtain its amplitude envelope representation. To improve robustness against noise and intra-user variability, a dual-smoothing strategy is employed using Gaussian smoothing kernels with different temporal scales. The fast Gaussian smoothing (illustrated in Figure 3a), obtained using a smaller standard deviation (), produces an envelope that preserves fine temporal transitions and captures transient blow onsets. In contrast, the slow Gaussian smoothing (illustrated in Figure 3b), obtained using a larger standard deviation (), yields a more stable envelope that captures long-term energy trends and sustained blow activity. A fused representation (illustrated in Figure 3c) combines the outputs of both Gaussian smoothing processes through mask intersection to retain consistently detected blow activity while suppressing spurious detections. Let denote the amplitude envelope of the preprocessed acoustic signal obtained after bandpass filtering and full-wave rectification. This process can be expressed as:
Figure 3. Gaussian smoothing modes: (a) fast (b) slow (c) fused
where and denote the fast- and slow-smoothed envelope representations, respectively, and denotes Gaussian smoothing with standard deviation , implemented as a convolution between the signal and a Gaussian kernel.
For each smoothed representation, segmentation is performed using a dynamic baseline gradient-tracking algorithm [38]. The first-order gradient of the smoothed envelope is computed to detect rapid energy transitions associated with blow onset activity. Once activated, a segment remains active as long as either the gradient magnitude remains sufficiently high or the signal energy exceeds an adaptive baseline estimated from preceding signal regions. To suppress false activations caused by transient fluctuations, a silence-reset mechanism is applied: if the gradient remains near zero and the signal energy stays below the adaptive baseline for a sustained duration, the active segment is terminated and the detector state is reset. The segmentation operator converts a smoothed continuous envelope into a binary sequence indicating active (blow) and inactive (non-blow) regions:
where and denote the binary segmentation sequences obtained from the fast and slow smoothed envelopes, respectively.
Following the segmentation stage, a post-processing refinement is applied to improve temporal consistency. Specifically, short inactive gaps between adjacent active regions are filled using a minimum-gap constraint, while segments shorter than a predefined duration threshold are removed. The above segmentation and post-processing pipeline is independently applied to both fast and slow smoothed representations, producing two candidate activity masks. The fused blow regions are obtained by intersecting both fast and slow smoothed representations using a logical AND operation (shown in Figure 3c), which reduces false positives while retaining consistently detected blow activity across both temporal resolutions. The fused blow regions () are defined as:
The fused blow regions () are used to extract the temporal feature set for the recording. Figure 4 illustrates the segmentation result for a single recording session, where the detected blow regions and inter-blow gaps are explicitly highlighted. Each contiguous blow region corresponds to an active blowing event, while the intervening non-active intervals represent inter-blow gaps, including the boundary intervals at the beginning and end of the recording. From , the number of blows () is obtained by counting the distinct blow regions, the duration of each blow () is computed from the temporal extent of its corresponding segment, and the inter-blow gaps () correspond to the non-active intervals between consecutive blow regions. In addition, the RMS amplitude within each blow region () is computed from the bandpass-filtered signal, the duty cycle () is computed as the proportion of the session duration occupied by blow activity, and the blow shape () is characterized by the rise and fall times around each blow’s peak amplitude.

Figure 4. Temporal feature extraction
The resulting per-session temporal feature representation (T), derived from , is formally defined as
where , , , , , and denote the blow count, blow durations, inter-blow gaps, RMS amplitudes, duty cycle, and blow shape described above, respectively.
Since T is extracted independently for each recording, it is not yet a comparison quantity. To obtain the temporal feature representation used in authentication, denoted , the enrolled and query sessions’ temporal features, and , are compared channel-wise: sequential channels (, , ) are compared using Dynamic Time Warping (DTW) alignment distance, while scalar channels (, , ) are compared using the absolute deviation between the enrolled and query values:
where denotes this per-channel comparison, and denotes the resulting raw temporal deviation for the current authentication attempt, analogous in form to the raw spectral similarity defined in § 3.3.
Feature Aggregation
To enhance authentication robustness under varying behavioral and environmental conditions, the proposed framework incorporates both spectral and temporal features described in § 3.3 and integrates them with LLM-based contextual reasoning. This enables adaptive context-aware authentication through the joint analysis of spectral and temporal biometric characteristics, rather than relying on a single similarity measure alone.
Since and are both already expressed as comparisons between the enrolled and query samples, they can be provided together as the input describing the authentication attempt:
where denotes the pair of comparison outputs for the current authentication attempt, each retained as a separate piece of evidence.
Since lightweight LLMs generally exhibit limited capability in directly interpreting raw numerical biometric features, a structured intermediate representation is introduced. A similarity computation function converts each channel of into a comparable similarity representation:
where P denotes the resulting unified similarity representation, consisting of normalized similarity scores. We evaluate two variants of : a rule-based mode, described in § 3.4, and a score-based mode, described in § 3.4.
Rule-Based The rule-based variant converts each channel deviation into a discrete similarity category {VERY_GOOD, GOOD, OKAY, WEAK, BAD} using fixed, globally-defined percentage-difference thresholds between the enrolled and attempted values, applied uniformly across all users. The embedding similarity channel is instead labeled directly from the cosine-vs-threshold margin, divided into the same five categories based on its position relative to the user’s decision threshold and the borderline band. Blow count is compared directly as an integer difference rather than a percentage, while the remaining channels (gap, duration, shape, amplitude, and mass) use symmetric percentage differences, with the shape channel calibrated to wider thresholds to reflect its larger typical variation.
Score-Based The score-based variant instead normalizes each channel of against the enrolled user’s own baseline reference statistics . By converting raw spectral and temporal deviations into a common, normalized scale, this feature aggregation stage allows the lightweight LLM to perform more effectively. Lightweight LLMs are generally less capable of directly interpreting raw numerical values. By expressing every feature channel as a bounded, uniformly interpretable similarity score, the aggregation stage removes this burden from the downstream reasoning module, allowing the LLM to focus on weighing and reasoning over the scores themselves rather than interpreting their raw numerical meaning.
LLM-Based Contextual Reasoning Engine
The proposed framework integrates a lightweight LLM as a high-level reasoning module for adaptive authentication. The LLM is responsible for interpreting structured behavioral representations and biometric similarity scores to perform context-aware authentication decisions that go beyond threshold-based verification. To enable effective reasoning, prompt engineering is employed to construct a structured and semantically consistent input prompt from the outputs of the feature aggregation module (cf. Section 3.4). The designed prompt incorporates the unified similarity representation derived from the similarity computation function , allowing the lightweight LLM to perform context-aware authentication reasoning.
Given the unified similarity representation , a prompt generation function is applied to construct the final LLM input prompt:
where Prompt represents the prompt generated.
The generated prompt Prompt is subsequently processed by the lightweight LLM to perform contextual authentication reasoning:
where denotes the lightweight LLM reasoning function and O represents the generated authentication response.
The generated response O consists of the authentication decision accompanied by a short natural language justification that explains the reasoning behind the decision based on the provided biometric and behavioral evidence.
Implementation Details
The proposed LABBA framework was implemented as a multi-stage behavioral biometric authentication system using Python. To collect blow-acoustic data, we develop a custom blow-based biometric authentication application for the Android platform. The application captures blow-acoustic signals through the device microphone at a sampling rate of 48 kHz. During acquisition, the raw audio signal is continuously buffered internally for the entire recording duration. Each recording session lasts for 5 seconds and contains the complete temporal acoustic waveform generated during the user’s blowing interaction. Blow-acoustic spectral features were extracted using the Gammatone[15] Python library for GFCC-based processing and CNN models implemented in TensorFlow, while blow-acoustic temporal features were derived using pure Python with NumPy and SciPy. Feature aggregation was implemented using custom Python logic. Per-user baseline reference statistics () are precomputed once from each user’s enrollment sessions. For each authentication attempt, the enrolled and query samples’ spectral and temporal features are compared to obtain and . Under rule-based mode, the similarity computation module maps and onto a fixed set of discrete categorical labels using globally-defined thresholds, requiring no per-user baseline precomputation. Under score-based mode, per-user baseline reference statistics () are instead precomputed once from each user’s enrollment sessions, and normalizes and against the corresponding user’s precomputed baseline to produce a continuous, bounded representation. In both cases, the resulting unified similarity representation is suitable for LLM-based reasoning and prompt construction. The resulting aggregated representation was transformed into prompts using prompt engineering and provided to a lightweight LLM, specifically Phi-3 Mini Instruct (Q4 quantized variant)[28], integrated via the llama-cpp-python[1] library, which performed contextual reasoning and generated the final authentication decision.
The source code and artifacts of LABBA are available at [5].
Evaluation
Experimental Setup
To collect blow-acoustic experimental data, we use the blow-acoustic application we developed. The extracted spectral and temporal blow-acoustic features were provided to a lightweight LLM model, specifically Phi-3 Mini Instruct (Q4 quantized version)[28], to generate authentication reasoning based on the captured blow-acoustic characteristics. The system was implemented using python and executed entirely on CPU-based processing without GPU acceleration. All experiments were conducted on a Dell XPS 15 9510 laptop equipped with an 11th Gen Intel Core i7-11800H processor (8 cores) and 32 GB RAM, running Windows.
Dataset
We conducted a user study involving 50 participants (40 males, 10 females) randomly selected from demographic groups ranging from 22 to 65 years of age. Each participant performed 10 sessions, with each session lasting 5 seconds, resulting in 500 total attempts. To ensure robustness and evaluate real-world applicability, the experiment was conducted under uncontrolled acoustic conditions, without a dedicated recording booth or acoustically treated room. Each participant completed all their sessions in their own chosen environment, reflecting the ambient acoustic variability typical of real-world smartphone usage, and they performed the experiment in both standing and sitting postural modes. To show the patterns and characteristics of the collected dataset, we illustrate the raw and refined blow-acoustic data of three participants in Figure 5 and 6, respectively.
Figure 5. Sample raw blow-acoustic data of (a) Participant 1 (b) Participant 2 (c) Participant 3
Accuracy Evaluation
To evaluate the accuracy of LABBA, recordings originating from the same participant were treated as genuine attempts, while recordings from different participants were treated as adversary attempts. Pairwise comparisons were performed across sessions to evaluate intra-user consistency and inter-user separability under varying environmental conditions. Effectiveness of LABBA was evaluated using standard biometric authentication metrics, including Accuracy, False Acceptance Rate (FAR), False Rejection Rate (FRR), and Equal Error Rate (EER). The EER is commonly defined as the operating point at which the FAR and FRR are equal. In this work, due to the discrete nature of the decision scores and the finite set of evaluation thresholds, we approximate the EER using the mean operating error rate, computed as: .
Figure 6. Sample refined blow-acoustic data of (a) Participant 1 (b) Participant 2 (c) Participant 3
Table 1 summarizes the biometric authentication accuracy of LABBA under configurations without and with LLM-assisted evaluation. The non-LLM baseline scheme achieves an accuracy of 99.09%, FAR of 0.13%, FRR of 43.64% and EER of 21.89%.
| Method | Gaussian | Accuracy (%) | FAR (%) | FRR (%) | EER (%) |
| Without LLM-Assisted | NA | 99.09 | 0.13 | 43.64 | 21.89 |
| With LLM-Assisted (Rule-Based) | Fast | 99.32 | 0.20 | 26.78 | 13.49 |
| With LLM-Assisted (Rule-Based) | Slow | 99.33 | 0.22 | 25.20 | 12.71 |
| With LLM-Assisted (Rule-Based) | Fused | 99.46 | 0.21 | 18.27 | 9.24 |
| With LLM-Assisted (Score-Based) | Fast | 99.25 | 0.58 | 14.22 | 7.40 |
| With LLM-Assisted (Score-Based) | Slow | 99.22 | 0.53 | 14.4 | 7.48 |
| With LLM-Assisted (Score-Based) | Fused | 99.38 | 0.38 | 13.8 | 7.09 |
Table 1. Accuracy results
Rule-Based Mode. Across all three Gaussian smoothing modes, rule-based mode substantially improves over the non-LLM baseline, with accuracy ranging from 99.32%–99.46%, FAR from 0.20%–0.22%, FRR from 18.27%–26.78%, and EER from 9.24%–13.49%. The Fused configuration achieves the best overall performance, with an accuracy of 99.46%, FAR of 0.21%, FRR of 18.27%, and EER of 9.24%, an improvement of 25.37 percentage points in FRR and 12.65 percentage points in EER over the non-LLM baseline, at the cost of only a 0.08 percentage point increase in FAR.
Score-Based Mode. Across all three Gaussian smoothing modes, score-based mode likewise substantially improves over the non-LLM baseline, with accuracy ranging from 99.22%–99.38%, FAR from 0.38%–0.58%, FRR from 13.80%–14.40%, and EER from 7.09%–7.48%. The Fused configuration again achieves the best overall performance, with an accuracy of 99.38%, FAR of 0.38%, FRR of 13.80%, and EER of 7.09%, an improvement of 29.84 percentage points in FRR and 14.80 percentage points in EER over the non-LLM baseline, at the cost of a 0.25 percentage point increase in FAR.
Comparing the two schemes’ best configurations, score-based mode yields a marginally lower accuracy and a modestly higher FAR than rule-based, but achieves a substantially lower FRR and EER, reducing FRR by roughly 4.47 percentage points and EER by roughly 2.15 percentage points relative to the best rule-based configuration. This indicates that the continuous, per-user-normalized representation used in score-based is more effective at correctly accepting genuine authentication attempts, at the cost of a slightly higher rate of impostor acceptance.
This suggests that score-based is preferable in usability-sensitive deployments where minimizing false rejections is a priority, while rule-based may be preferable where minimizing false acceptance is the primary concern. Across both schemes, the fused configuration consistently achieves the lowest EER, indicating that combining multi-scale Gaussian smoothing improves the robustness and discriminative capability of the extracted behavioral representation by jointly capturing both transient and sustained blowing patterns. While both LLM-assisted schemes yield substantial improvements over the non-LLM baseline, they do not eliminate authentication errors entirely; false acceptances and false rejections still occur under certain conditions, as reflected in the FAR and FRR values above. Beyond these quantitative improvements, the LLM-assisted approach additionally provides a natural language justification for each authentication decision, offering a level of interpretability that is not available in the conventional threshold-based scheme.
Illustrating the LLM Output with Examples
Table 3 presents representative examples of LLM-assisted biometric decision-making under the two evaluated modes, rule-based and score-based modes, described in § 3.4 and § 3.4, respectively. For each scheme, one genuine-accept and one correct-reject case are shown, allowing the LLM’s reasoning to be examined along two axes: how it behaves when accepting versus rejecting an attempt, and how that reasoning differs depending on whether it is given discrete categorical labels or continuous normalized scores as input. The first two examples correspond to rule-based mode, where each channel is expressed as one of five discrete labels (VERY_GOOD, GOOD, OKAY, WEAK, BAD), while the last two examples correspond to score-based mode, where each channel is instead expressed as a continuous similarity value on a 0–100 scale.
| Input Prompt | LLM Response | Expected |
| TASK: Decide ACCEPT or REJECT for a blow pattern authentication attempt. - embedding similarity: VERY_GOOD - blow count similarity: VERY_GOOD - gap similarity: OKAY - duration similarity: VERY_GOOD - shape similarity: OKAY - amplitude similarity: VERY_GOOD - mass similarity: VERY_GOOD Return ONLY: Decision: ACCEPT or REJECT Confidence: 0-100 Reason: Describe Accept/Reject Decision | Decision: ACCEPT Confidence: 95 Reason: The blow pattern authentication attempt has shown very good similarity in embedding, blow count, duration, amplitude, and mass. Although the gap and shape similarities are only okay, the overall high similarity in critical aspects of the blow pattern leads to a decision of ACCEPT. The gap and shape similarities are not significantly different to warrant a rejection. | ACCEPT |
| TASK: Decide ACCEPT or REJECT for a blow pattern authentication attempt. - embedding similarity: BAD - blow count similarity: VERY_GOOD - gap similarity: BAD - duration similarity: OKAY - shape similarity: VERY_GOOD - amplitude similarity: OKAY - mass similarity: OKAY Return ONLY: Decision: ACCEPT or REJECT Confidence: 0-100 Reason: Describe Accept/Reject Decision | Decision: REJECT Confidence: 95 Reason: The blow pattern authentication attempt has a significant issue with the embedding similarity, which is marked as BAD. This indicates that the audio pattern does not match the expected embedding, making it highly unlikely that the attempt is genuine. Although the blow count, shape, amplitude, and mass similarities are good, the poor embedding similarity is a critical factor that leads to the rejection of the authentication attempt. | REJECT |
| TASK: Decide ACCEPT or REJECT for this blow pattern authentication attempt. SIMILARITY SCORES (0-100 each; 100 = matches the enrolled pattern exactly, 0 = no resemblance; each already accounts for this user’s own normal variability): - embedding similarity: 100/100 - blow count similarity: 100/100 - gap pattern similarity: 90/100 - duration pattern similarity: 75/100 - shape pattern similarity: 80/100 - amplitude (loudness) similarity: 95/100 - blow "mass" (total effort) similarity: 60/100 Return ONLY: Decision: ACCEPT or REJECT Confidence: 0-100 Reason: Describe Accept/Reject Decision | Decision : ACCEPT Confidence: 85 Reason : The blow pattern authentication attempt has high similarity scores in embedding, blow count, gap pattern, and amplitude similarity, which indicates a strong resemblance to the enrolled pattern. Although there are some differences in duration and shape pattern similarity, the overall similarity is high, and the blow "mass" similarity is also relatively close. Therefore, the authentication attempt can be accepted with a confidence level of 85. | ACCEPT |
| TASK: Decide ACCEPT or REJECT for this blow pattern authentication attempt. SIMILARITY SCORES (0-100 each; 100 = matches the enrolled pattern exactly, 0 = no resemblance; each already accounts for this user’s own normal variability): - embedding similarity: 0/100 - blow count similarity: 100/100 - gap pattern similarity: 10/100 - duration pattern similarity: 15/100 - shape pattern similarity: 95/100 - amplitude (loudness) similarity: 40/100 - blow "mass" (total effort) similarity: 5/100 Return ONLY: Decision: ACCEPT or REJECT Confidence: 0-100 Reason: Describe Accept/Reject Decision | Decision : REJECT Confidence: 70 Reason : While the blow count and shape pattern similarities are high, indicating a strong resemblance to the enrolled pattern, the low blow "mass" similarity and gap pattern similarity suggest significant differences in the blow pattern. The overall similarity scores are not high enough to confidently accept the authentication attempt. | REJECT |
Table 3. LLM evaluation examples (rule-based and score-based)
The table also presents the evaluation result generated by the lightweight LLM based on the corresponding input prompt. For each example, the LLM returns a single decision, ACCEPT or REJECT, accompanied by a natural language reason justifying that decision. This illustrates the interpretability framework provides alongside its authentication outcome, as the reasoning behind each decision remains inspectable regardless of which mode, rule-based or score-based, is used to construct the prompt.
Attack Resilience Evaluation
The proposed LLM-assisted blow-acoustic behavioral biometric authentication framework demonstrates inherent resilience against several common biometric attacks due to the dynamic and context-dependent nature of blow-acoustic interactions. The framework is naturally resistant to mimicry attacks because personalized blow-acoustic patterns contain complex temporal, spectral, and behavioral characteristics that are difficult to accurately imitate.
The proposed approach is also resilient against biometric template compromise attacks since it does not rely on storing conventional fixed biometric templates. Instead, authentication is performed using dynamically generated and processed behavioral embeddings. In addition, the framework exhibits natural robustness against replay attacks because capturing and realistically replaying blow-acoustic interactions is highly challenging. Furthermore, modern liveness detection techniques [9] can be easily integrated to further strengthen replay attack resistance.
It is also difficult to attack the framework using AI-generated synthetic acoustic sounds. Although synthetic sounds can imitate certain acoustic characteristics, they often fail to reproduce the rich contextual and behavioral information required by the LLM-based reasoning engine. As a result, the proposed framework evaluates not only signal similarity but also contextual consistency and behavioral coherence during authentication.
Compared to general-purpose LLM models, the proposed framework is relatively less vulnerable to prompt injection and jailbreaking attacks due to its constrained reasoning architecture. The LLM operates exclusively on structured behavioral biometric and contextual inputs generated by trusted system components, rather than unrestricted natural language prompts, which reduces the attack surface. This design also mitigates a recently demonstrated class of physical-layer attacks against speech-driven LLMs, in which an adversary covertly injects jailbreak prompts as inaudible near-ultrasonic audio recovered only after microphone nonlinearity, achieving high success rates specifically against systems that process raw audio directly [25]. Because LABBA’s LLM never receives raw acoustic input, only structured similarity representations produced by the similarity computation function , this covert acoustic injection channel does not apply to our framework. In addition, the LLM is restricted to bounded contextual authentication reasoning within controlled boundaries. Nevertheless, additional strategies, e.g., trusted training, constrained fine-tuning, and contextual consistency validation, may still be required to further enhance the robustness against adversarial manipulation.
Usability Evaluation
The proposed LLM-assisted blow-acoustic behavioral biometric authentication framework demonstrates strong usability in terms of ease of use, non-invasiveness, and contextual robustness. The authentication process requires only a few natural blowing interactions toward the smartphone microphone, making the system simple and convenient for users. In addition, the approach is non-invasive and contactless, as it does not require physical touch interaction with the smartphone during authentication.
The proposed framework also exhibits strong robustness under dynamic behavioral and environmental conditions. Experimental evaluations were conducted under diverse contexts, including sitting and standing postural positions, varying blowing intensities and patterns, and different phone interaction dynamics. Despite these variations, the authentication performance remained highly stable because the LLM-based reasoning engine adaptively interprets natural behavioral and contextual variability.
Furthermore, the system was evaluated under real-world acoustic conditions, involving quiet and noisy environments. The authentication performance remained reliable due to the use of band-pass filtering and Gaussian smoothing techniques, which effectively suppress environmental noise and unwanted frequency components prior to feature extraction, which is intended to support reliable performance across varying acoustic environments.
We acknowledge that LLM outputs are stochastic and may occasionally be unfaithful to the input (hallucination), which is a practical concern for an authentication decision. A systematic study of decision consistency across repeated runs and prompt variations is left as future work.
Overall, the proposed framework provides a practical, user-friendly, and context-aware authentication solution suitable for real-world deployment environments.
Performance Evaluation
Table 2 summarizes the runtime performance of the proposed protocol, including spectral feature extraction, temporal feature extraction under different Gaussian smoothing configurations (Fast, Slow, and Fused), and LLM-assisted evaluation. In particular, the Fast configuration provides lower computational cost due to smaller Gaussian smoothing kernels, whereas the Slow configuration incurs higher runtime as a result of larger smoothing windows. The Fused configuration further increases computational cost as it jointly executes both Fast and Slow temporal processing pipelines, resulting in additional processing overhead. The table also shows that LLM evaluation under score-based mode takes noticeably longer than under rule-based mode (11.604s vs. 8.237s). This is primarily attributable to prompt length: the score-based prompt includes an additional explanatory clause describing the 0–100 similarity scale, and represents each channel as a numeric ratio (e.g., 100/100) rather than a single categorical label (e.g., VERY_GOOD), resulting in a longer input sequence for the LLM to process. This reflects a further trade-off beyond the accuracy-usability trade-off discussed in § 4.3: the more precise, continuous representation used in score-based mode incurs additional runtime overhead, but corresponds to the mode that also achieves the lower FRR reported in Table 1, indicating that improved usability through more precise prompting comes at the cost of increased inference latency. Feature aggregation and prompt generation introduce negligible computational overhead and do not significantly contribute to the overall execution time.
| Method | Time (s) |
| Spectral Feature Extraction | 0.589 |
| Temporal (Fast-Smoothed) Feature Extraction | 0.486 |
| Temporal (Slow-Smoothed) Feature Extraction | 2.350 |
| Temporal (Fused-Smoothed) Feature Extraction | 2.832 |
| Feature Aggregation | |
| Prompt Generation | |
| LLM Evaluation (Rule-Based) | 8.237 |
| LLM Evaluation (Score-Based) | 11.604 |
Table 2. Runtime performance results
The reported execution times represent the average runtime per authentication session under the experimental setup described previously. Although LLM-assisted evaluation introduces the highest inference latency, the current implementation operates entirely on CPU without GPU acceleration or deployment-level optimization. As such, the reported runtime reflects the computational characteristics of a research prototype and may be further reduced through hardware acceleration and optimized inference strategies. We acknowledge that the current end-to-end runtime, ranging from 8.237s to 11.604s depending on the scoring mode, is too slow for real-time authentication deployment regardless of which mode is used. However, the primary focus of this work is to investigate whether LLM-based contextual reasoning can meaningfully assist behavioral biometric authentication, rather than to optimize for deployment-level efficiency at this stage. Addressing runtime and usability constraints through hardware acceleration, model optimization, and inference-efficient deployment strategies is left as future work.
Related work
Behavioral Biometrics
Behavioral biometrics [14, 13, 21] have gained increasing attention due to their continuous and unobtrusive authentication capabilities and their potential to enhance resilience against spoofing attacks by leveraging distinctive behavioral characteristics. Closer to our own modality, blowing-based acoustic patterns have recently been proposed as a behavioral biometric [18], using DTW for fixed similarity-based verification. A follow-up work [9] extended this modality using GFCC-based spectral features with a CNN encoder, similarly relying on fixed cosine similarity for the final decision. Both works establish blow-acoustic signals as a viable behavioral biometric modality but rely on conventional, fixed similarity-based decision-making without dynamic, contextual reasoning, motivating our integration of LLM-assisted reasoning into such modality.
LLM-Assisted Biometrics
Recently, there has been increasing interest in integrating LLMs into biometric systems for tasks such as biometric detection [27, 2, 32] and recognition / authentication[11, 19, 12, 35, 44, 30] across different biometric modalities. This growing intersection between foundation models, especially LLMs, and biometrics has recently been surveyed in the literature [31].
Mehta et al. [27] explored the use of keystroke dynamics as a behavioral biometric for detecting academic dishonesty. Their work benchmarked LLaMA-3.3-70B-Instruct for identifying LLM-assisted cheating, demonstrating the potential of LLMs for behavioral biometric-based attack detection. Almehmadi [2] introduced a framework for detecting coercion as a security state using LLM-assisted multimodal biometrics. The approach integrates behavioral signals such as keystroke dynamics and mouse movements, physiological signals including heart-rate variability, and contextual user commands to infer potential coercion scenarios. Piland et al. [32] extended LLM-based biometric reasoning to visual modalities by investigating iris presentation attack detection (PAD) using multimodal LLMs. Their results show that models such as Gemini and Llama-Vision can leverage pretrained visual representations and human salience-guided prompting to distinguish bona fide from attack samples without task-specific training. These studies [27, 2, 32] reflect a growing trend in leveraging LLMs for reasoning and decision-making over multimodal biometric signals.
Deandres-Tame et al. [11] investigated the effectiveness of LLMs, particularly ChatGPT, in face biometric applications, including face recognition, soft biometric attribute analysis, and explainable biometric reasoning. Similarly, Hassanpour et al. [19] evaluated ChatGPT for biometric-related tasks such as face recognition, gender detection, and age estimation. Their work demonstrated that LLMs are capable of generating human-readable explanations for biometric decisions by describing distinguishing facial characteristics and auxiliary biometric traits. Similarly, Farmanifard et al. [12] explored the use of LLMs, including ChatGPT and Gemini, for iris biometric authentication tasks. Furthermore, Rastogi [35] investigated the application of LLMs, particularly ChatGPT-4, in vascular biometric recognition. The study
Description of Notations: N/A: not applicable due to ChatGPT4 and Gemini do not publicly report their parameters; ✓: achieved; ×: not achieved; *: based on the ROC-AUC metric; †: rule-based; ‡: score-based; –: not reported or available
showed that ChatGPT-4 achieved strong adaptability and accuracy in vein identification tasks across multiple scenarios. These studies [12, 35] collectively demonstrate that LLMs can exhibit notable adaptability in biometric recognition when guided by appropriate prompt engineering strategies. However, these approaches [11, 12, 35] primarily rely on large-scale proprietary LLMs that require cloud-based API access, resulting in significant computational and operational costs that limit their deployment in resource-constrained environments.
In the audio domain, Nam et al. [30] proposed SpeakerLLM, a speaker-specialized audio-LLM for speaker verification reasoning, achieving 97.12% verdict accuracy on VoxCeleb1-O. Like our approach, it uses a lightweight LLM (Qwen2.5-1.5B-Instruct). However, it requires training a custom audio tokenizer so the LLM can understand speaker embeddings directly. Our approach needs no such training: biometric evidence is converted into plain text before reasoning, so an off-the-shelf LLM can be used directly through prompt engineering alone.
Comparison of Related Works
To the best of our knowledge, there is limited existing work that directly investigates LLM-based behavioral biometrics. Therefore, we identify the most closely related available studies in behavioral biometric authentication and compare their approaches with our proposed work, LABBA, as summarized in Table 4. Most existing LLM-assisted biometric studies primarily focus on physiological biometrics, particularly face, iris, and vascular image recognition and analysis. In contrast, the use of LLMs in behavioral biometrics has largely been investigated for attack detection and malicious activity identification rather than for user authentication. LABBA explores a novel direction by introducing LLM-assisted authentication into the behavioral biometrics domain, specifically for phone-blowing acoustic behavioral biometrics. Therefore, it demonstrates the feasibility of leveraging LLM-based reasoning for behavioral biometric authentication and extends the applications of LLMs beyond the conventional physiological biometric scenarios.
| Related Work | Modality | Acc (%) | EER (%) | Usage | Model | LLM Model Parameters | Context-Aware | LLM-Reasoning-Based Decision |
| Mehta et al. [27] | Keystroke Dynamics | – | – | Attack Detection | LLaMA-3.3-Instruct | 70B | × | ✓ |
| Almehmadi [2] | Keystroke Dynamics, Mouse-Movement, Heart Rate | – | 0* | Attack Detection | DeepSeek-R1-Distill-Qwen | 32B | ✓ | ✓ |
| Piland et al. [32] | Iris Biometrics | – | – | Attack Detection | Gemini 2.5 Pro Llama 3.2-Vision | 90B | ✓ | ✓ |
| Deandres-Tame et al. [11] | Face Biometrics | 94 | – | Face Verification | GPT-4V | N/A | × | ✓ |
| Hassanpour et al. [19] | Face Biometrics | 95.15 | – | Face Verification | GPT-4V | N/A | × | ✓ |
| Farmanifard et al. [12] | Iris Biometrics | – | – | Authentication | GPT-4V, Gemini | N/A | × | ✓ |
| Rastogi et al. [35] | Vascular Biometrics | – | – | Authentication | GPT-4V, Gemini | N/A | × | ✓ |
| Zerroug et al. [44] | Iris Images | – | – | Biometric Enrichment | BERT, GPT-4, LLaMa | N/A | × | × |
| Nam et al. [30] | Voice/Speaker Biometrics | 97.12 | – | Speaker Verification | Qwen2.5-1.5B-Instruct | 1.5B | ✓ | ✓ |
| Halim et al. [18] | Acoustic-based Biometrics | 99.35 | 9.74 | Authentication | – | – | × | × |
| LABBA (ours) | Acoustic-based Biometrics | 99.46†/99.38‡ | 9.24†/7.09‡ | Authentication | Phi-3 Mini Instruct | 3.8B | ✓ | ✓ |
Table 4. Comparison of related works
In addition, although Halim et al. [18] is not an LLM-based work, we include it in our comparison because it employs a closely related blow-acoustic biometric modality. While their approach does not incorporate reasoning-based authentication decisions, it reports an accuracy of 99.35%, which is slightly lower than the accuracy achieved by LABBA (99.46% with the rule-based mode and 99.38% with the score-based mode). The reported EER of Halim et al. (9.74%) is also higher than that of LABBA (9.24% with the rule-based mode and 7.09% with the score-based mode). Moreover, we note that the reported accuracy in Halim et al. [18] is obtained using a session-to-user comparison based on KNN, where each probe session is compared only with the nearest sessions among the available sessions of a user, rather than being evaluated through exhaustive pairwise (session-to-session) comparisons. Such an evaluation may yield higher accuracy because in KNN, each probe session is compared against only the most similar sessions rather than directly against each individual biometric session. In contrast, LABBA adopts a pairwise evaluation, in which each biometric session is directly compared with other sessions. This provides a more fine-grained evaluation of the system’s ability to distinguish genuine and impostor biometric sessions.
Conclusion
In this work, we proposed an LLM-assisted behavioral biometric authentication framework that performs authentication through contextual trust evaluation rather than relying solely on a single fixed similarity threshold. The proposed framework leverages an LLM as a contextual authentication reasoner capable of integrating spectral and temporal biometric similarity descriptors to support adaptive authentication decisions. The framework is evaluated using a novel phone-blowing behavioral biometric technique that captures rich environmental and behavioral contextual information.
Experimental results demonstrate that the proposed approach improves authentication flexibility and reduces the false rejection rate while maintaining a low false acceptance rate, achieving an overall accuracy of 99.46% and 99.38% under rule-based and score-based modes, respectively. The findings establish a new direction for behavioral biometric authentication by transforming conventional signal-based verification into an intelligent cognitive authentication framework capable of adaptive, context-aware, and semantically informed security decision-making.
Despite the promising results, several directions remain for future research. First, the current implementation operates entirely on CPU-based execution, and future work may explore GPU acceleration and cloud-based deployment strategies to improve inference efficiency and scalability. Second, alternative LLM models with varying model sizes and computational requirements may be investigated to further optimize the trade-off between computational cost and authentication performance. Finally, future studies should examine security and privacy-preserving mechanisms for LLM-assisted biometric authentication, including secure feature handling, privacy-aware inference, and resistance against adversarial manipulation.
Open Science
The source code and artifacts of LABBA are available at [5].
References
- [1]abetlen: llama-cpp-python: Python bindings for llama.cpp. https://github.com/abetlen/llama-cpp-python (2024), accessed: 2026-05-27
- [2]Almehmadi, A.: Cyber coercion detection using llm-assisted multimodal biometric system. Applied Sciences 15(19), 10658 (2025)DOI
- [3]Alsaadi, I.M.: Physiological biometric authentication systems, advantages, disadvantages and future development: A review. International Journal of Scientific & Technology Research 4(12), 285–289 (2015)
- [4]Ang, K.W., Chekole, E.G., Zhou, J.: Unveiling the covert vulnerabilities in multi-factor authentication protocols: A systematic review and security analysis. ACM Comput. Surv. 57(11) (Jun 2025). https://doi.org/10.1145/3734864, https://doi.org/10.1145/3734864
- [5]Authors, L.: LABBA Dataset (2026), https://github.com/dtc-project/LABBA, to be treated confidentially until it is made publicly available.
- [6]Avola, D., Cinque, L., De Marsico, M., Fagioli, A., Foresti, G.L., Mancini, M., Mecca, A.: Signal enhancement and efficient dtw-based comparison for wearable gait recognition. Computers & Security 137, 103643 (2024)DOI
- [7]Ayeswarya, S., Singh, K.: Enhancing security and usability with context aware multi-biometric fusion for continuous user authentication. Scientific Reports 15(1), 30627 (2025)DOI
- [8]Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., Wang, Y., Ye, W., Zhang, Y., Chang, Y., Yu, P.S., Yang, Q., Xie, X.: A survey on evaluation of large language models. ACM Trans. Intell. Syst. Technol. 15(3) (Mar 2024). https://doi.org/10.1145/3641289, https://doi.org/10.1145/3641289
- [9]Chekole, E.G., Halim, H., Reijsbergen, D., Zhou, J.: Blowlive: Blow-based multi-factor biometrics with liveness detection and revocability. In: European Symposium on Research in Computer Security. pp. 585–605. Springer (2026)arxiv.org/abs/2606.25998
- [10]Chimuco, F.T., Sequeiros, J.B., Lopes, C.G., Simões, T.M., Freire, M.M., Inácio, P.R.: Secure cloud-based mobile apps: attack taxonomy, requirements, mechanisms, tests and automation. International Journal of Information Security 22(4), 833–867 (2023)DOI
- [11]Deandres-Tame, I., Tolosana, R., Vera-Rodríguez, R., Morales, A., Fierrez, J., Ortega-Garcia, J.: How good is chatgpt at face biometrics? a first look into recognition, soft biometrics, and explainability. IEEE Access 12, 34390–34401 (2024)
- [12]Farmanifard, P., Ross, A.: Chatgpt meets iris biometrics. In: 2024 IEEE International Joint Conference on Biometrics (IJCB). pp. 1–10. IEEE (2024)
- [13]Fereidooni, H., König, J., Rieger, P., Chilese, M., Gö kbakan, B., Finke, M., Dmitrienko, A., Sadeghi, A.R.: AuthentiSense: A scalable behavioral biometrics authentication scheme using few-shot learning for mobile platforms. In: Network and Distributed System Security (NDSS) Symposium (2023). https://doi.org/10.14722/ndss.2023.23194
- [14]Finnegan, O., White III, J., Armstrong, B., Adams, E., Burkart, S., Beets, M., Nelakuditi, S., Willis, E., Von Klinggraeff, L., Parker, H., et al.: The utility of behavioral biometrics in user authentication and demographic characteristic detection: a scoping review. Systematic Reviews 13(1), 61 (2024)
- [15]Gammatone Developers: Gammatone: Python package for auditory filterbank processing. https://pypi.org/project/Gammatone/ (2024), accessed: 2026-05-27
- [16]Gheorghian, M.G., Socotar, D.C., Ciobanu, R.C.: Study on anomaly detection in biometric data: Methods, challenges, and applications in information security. In: 2024 IEEE International Conference on Electrical and Power Engineering (EPEi). pp. 261–264. IEEE (2024)DOI
- [17]Habib, S., Khan, H., Hamilton-Wright, A., Hengartner, U.: Revisiting the security of biometric authentication systems against statistical attacks. ACM Transactions on Privacy and Security 26(2), 1–30 (2023)DOI
- [18]Halim, H., Chekole, E.G., Reijsbergen, D., Zhou, J.: Blowprint: Blow-based multi-factor biometrics for smartphone user authentication. In: European Symposium on Research in Computer Security. pp. 169–189. Springer (2025)arxiv.org/abs/2507.04126
- [19]Hassanpour, A., Kowsari, Y., Otroshi Shahreza, H., Yang, B., Marcel, S.: ChatGPT and biometrics: An assessment of face recognition, gender detection, and age estimation capabilities. In: IEEE International Conference on Image Processing (ICIP) (2024). https://doi.org/10.1109/ICIP51287.2024.10647924
- [20]Jain, A.K., Ross, A., Uludag, U.: Biometric template security: Challenges and solutions. In: 2005 13th European signal processing conference. IEEE (2005)DOI
- [21]Khan, S., Devlen, C., Manno, M., Hou, D.: Mouse dynamics behavioral biometrics: A survey. ACM Computing Surveys 56(6), 1–33 (2024)DOI
- [22]Khine, K.K., Su, C.: Acoustic scene classification using deep c-rnn based on log mel spectrogram and gammatone frequency cepstral coefficients features. In: 2024 3rd International Conference on Artificial Intelligence For Internet of Things (AIIoT). pp. 1–6 (2024). https://doi.org/10.1109/AIIoT58432.2024.10574640
- [23]Kokal, S., Pryor, L., Dave, R.: Exploration of machine learning classification models used for behavioral biometrics authentication. In: Proceedings of the 2022 8th International Conference on Computer Technology Applications. pp. 176–182 (2022) Howard Halim, Eyasu Getahun Chekole, and Jianying Zhou
- [24]Li, L., Chen, C., Pan, L., Tai, Y., Zhang, J., Xiang, Y.: Hiding your signals: A security analysis of ppg-based biometric authentication. In: European Symposium on Research in Computer Security. pp. 183–202. Springer (2023)arxiv.org/abs/2207.04434
- [25]Ling, Z., Hu, P., Gao, X., Ma, X., Zhou, M., Feng, J., Lu, S., Zhang, D., Zhu, B.B.: Sirens’ Whisper: Inaudible near-ultrasonic jailbreaks of speech-driven LLMs. In: 35th USENIX Security Symposium (USENIX Security 26) (2026)arxiv.org/abs/2603.13847
- [26]Madavarapu, J.B., Mittal, M., Salagrama, S., Adnan, M.M., Rana, A., Yadav, K.: Behavioral biometrics authentication systems: leveraging machine learning for enhanced cybersecurity. In: 2024 international conference on communication, computer sciences and engineering (ICC3SE). pp. 1478–1483. IEEE (2024)DOI
- [27]Mehta, A., Kumar, R., Singla, A., Bisht, K., Singla, Y.K., Shah, R.R.: Detecting llm-assisted academic dishonesty using keystroke dynamics. IEEE Transactions on Biometrics, Behavior, and Identity Science (2026)arxiv.org/abs/2511.12468
- [28]Microsoft: Phi-3-mini-4k-instruct-gguf. Hugging Face (2024), https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-ggufDOI
- [29]Nair, V., Hinton, G.E.: Rectified linear units improve restricted Boltzmann machines. In: Proceedings of the 27th International Conference on Machine Learning (ICML). pp. 807–814 (2010)
- [30]Nam, K., Heo, J., Bae, S., Yu, H.J., Chung, J.S.: Speakerllm: A speaker-specialized audio-llm for speaker understanding and verification reasoning. arXiv preprint arXiv:2605.15044 (2026)arxiv.org/abs/2605.15044
- [31]Otroshi Shahreza, H., Marcel, S.: Foundation models and biometrics: A survey and outlook. IEEE Transactions on Information Forensics and Security 20, 9113–9138 (2025). https://doi.org/10.1109/TIFS.2025.3602233
- [32]Piland, J., Dowling, B., Sweet, C., Czajka, A.: Generalist multimodal llms gain biometric expertise via human salience for iris presentation attack detection. IEEE Access (2026)DOI
- [33]Pisani, P.H.: A study on user-specific threshold configuration for keystroke dynamics in the context of adaptive biometric systems. In: Simpósio Brasileiro de Segurança da Informação e de Sistemas Computacionais (SBSeg). pp. 725–731. SBC (2024)DOI
- [34]Qin, D., Amariucai, G., Qiao, D., Guan, Y.: Improving behavior based authentication against adversarial attack using xai. arXiv preprint arXiv:2402.16430 (2024)arxiv.org/abs/2402.16430
- [35]Rastogi, S.: Vascular biometrics meets llms. Authorea Preprints (2025)DOI
- [36]Shadman, R., Wahab, A.A., Manno, M., Lukaszewski, M., Hou, D., Hussain, F.: Keystroke dynamics: Concepts, techniques, and applications. ACM Computing Surveys 57(11), 1–35 (2025)DOI
- [37]Sharma, S.B., Dhall, I., Nayak, S.R., Chatterjee, P.: Reliable biometric authentication with privacy protection. In: Advances in Communication, Devices and Networking: Proceedings of ICDN 2021, pp. 233–249. Springer (2022)DOI
- [38]Sharma, S.D., Rajawat, K.: Optimized gradient tracking for decentralized online learning. IEEE Transactions on Signal Processing 72, 1443–1459 (2024)arxiv.org/abs/2306.06375
- [39]Solano, J., Camacho, L., Correa, A., Deiro, C., Vargas, J., Ochoa, M.: Combining behavioral biometrics and session context analytics to enhance risk-based static authentication in web applications. International Journal of Information Security 20(2), 181–197 (2021)DOI
- [40]Solano, J., Lopez, C., Rivera, E., Castelblanco, A., Tengana, L., Ochoa, M.: Scrap: synthetically composed replay attacks vs. adversarial machine learning attacks against mouse-based biometric authentication. In: Proceedings of the 13th ACM Workshop on Artificial Intelligence and Security. pp. 37–47 (2020)
- [41]Souza, L.M.d., Guido, R.C., Contreras, R.C., Viana, M.S., Bongarti, M.A.d.S.: Improving voice spoofing detection through extensive analysis of multispectral feature reduction. Sensors 25(15), 4821 (2025)DOI
- [42]Subash, A., Song, I., Lee, I., Lee, K.: Adaptability of current keystroke and mouse behavioral biometric systems: A survey. Computers & Security 160, 104731 (2026). https://doi.org/10.1016/j.cose.2025.104731, https://www.sciencedirect.com/science/article/pii/S0167404825004201
- [43]Wang, X., Huang, Y., Wang, Y., Luo, X., Guo, K., Zhou, Y., Zhang, X.: AdaReasoner: Adaptive reasoning enables more flexible thinking. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 38, pp. 75737–75764 (2025)DOI
- [44]Zerroug, M., Radouane, M.: Automatic enrichment of biometric databases with metadata using llms. In: 2025 International Conference on Intelligent Systems: Theories and Applications (SITA). pp. 1–7. IEEE (2025)DOI


