Evidence-Gated Security and Performance Evaluation of Hash-Chain-Audited Biometric Verification for Smart Parking

Urban parking management is increasingly digitized through mobile applications, sensing, electronic payment, and data-driven platforms. However, many systems optimize space availability, payment, or traffic flow without treating field-officer identity and verification-record integrity as linked security-engineering requirements. Recent smart-parking reviews describe the transition toward integrated services [1], while Indonesian implementations demonstrate the operational relevance of roadside supervision and public accountability [2]. In the Medan use case, a mobile face-verification service is intended to help distinguish registered parking officers from unknown persons, but the resulting decision must also remain auditable after the transaction.

Biometric verification and audit logging are treated here as complementary controls with different responsibilities. The recognition subsystem estimates identity membership, whereas the audit subsystem checks the internal consistency and linkage of stored decisions under a trusted verifier. The ledger does not improve recognition accuracy, and the recognition model does not prevent later record modification. Unkeyed hash linkage alone does not authenticate the stored history against a writer who can replace records and recompute subsequent hashes. The security interpretation therefore depends on the stated adversary, trust assumptions, and evidence boundary.

This distinction is necessary because the available evidence is heterogeneous. Recognition and ledger timing were collected in separate experiments; the biometric data contain repeated placements of a small local identity set; and the ledger benchmark is a sparse fractional allocation rather than a balanced factorial experiment. Moreover, several archived variables were later found to be unsuitable as independent optimization outcomes. A safety- and security-oriented analysis should expose these limitations rather than convert them into deployment claims.

Accordingly, this study addresses three questions: (1) what within-dataset information about route utilization, errors, and recognition latency is supported by the repeated biometric placements; (2) how do the configured ledger mechanism and worker count relate to directly measured append latency, including interaction sensitivity and exploratory scores on a reused validation subset; and (3) which isolated record-integrity inconsistencies were detected by the trusted verifier, which structural operations were prevented by database constraints, and which adversarial conditions remain untested?

The paper contributes an explicit evidence and threat boundary, a reproducible specification of the quality-aware calibrated open-set face-identification workload (QAC-OSFI), an audit of measurement provenance and experimental allocation, an exploratory interaction-aware append-latency analysis, and a mechanism-specific account of integrity outcomes. The scope is limited to observed consistency checks and measured append-performance associations. Neither independent predictive validation nor resistance to a consistently rewritten hash-chain history is established.

3.1 Evidence and threat boundary

Figure 1 depicts the actual evidence boundary. Recognition replay and audit benchmarking were conducted as separate experiments. Stored recognition timings were combined offline with isolated ledger timings to form a composite benchmark value; this value is not a synchronously measured mobile-camera-to-audit transaction. The evidence-admissibility gate retains only outcomes whose provenance supports the stated analysis. For record integrity, the evaluated evidence is detection of unrehashed inconsistencies by a trusted verifier, not authentication of history against a database writer who recomputes the affected hash chain. Table 1 summarizes the evaluated threats, implemented controls, observed evidence, and explicit boundaries.

1.png

Figure 1. Measurement provenance, security boundary, and evidence-admissibility gate

Table 1. Threat model, implemented controls, and tested boundaries

3.2 Frozen quality-aware calibrated open-set face-identification workload specification

MobileFaceNet is the fast backbone and AdaFace IR-50 is the expert backbone. All embeddings are L2-normalized. For each aligned enrollment image, the scalar quality score combines YuNet detection confidence $d$, normalized Laplacian sharpness $s$, and normalized face-size ratio $r$. The result is clipped to [0, 1] and rounded to four decimal places; a detected face is rejected before embedding when either bounding-box side is below 60 pixels.

$q=\operatorname{clip}(0.45 d+0.35 s+0.20 r, 0,1)$         (1)

Enrollment images are combined using quality-weighted prototypes with the frozen temperature $\beta=2$. Only ENROLLMENT images contribute to a prototype; VALIDATION and UNKNOWN_VALIDATION data are used for calibration and thresholds, whereas PROBE and UNKNOWN_PROBE data are reserved for reporting.

$w_{i j}=\frac{\exp \left(2 q_{i j}\right)}{\sum_k \exp \left(2 q_{i k}\right)}, p_i=\operatorname{normalize}\left(\sum_j w_{i j} e_{i j}\right)$        (2)

The fast-stage calibrator is deterministic L2-regularized logistic regression over the standardized feature vector $z_{\mathrm{F}}=$ $\left[s_{(1)}, \Delta, q, \log \left(1+n_{i^*}\right)\right]$, where $s_{(1)}$ is the top cosine score, $\Delta$ is the top-1/top-2 margin, and $n_{i^*}$ is the enrollment-support count of the leading candidate. Standardization uses the population means and standard deviation with a $10^{-6}$ floor. Full-batch gradient descent is run for 800 iterations with a learning rate of 0.05 and an L2 coefficient of 0.001.

$\pi_F=\sigma\left(w_F{ }^T z_F+b_F\right)$          (3)

For seeds 42, 101, 202, 303, and 404, the fast thresholds are 0.156261, 0.236002, 0.126669, 0.188091, and 0.122805, respectively. Each threshold equals the maximum unknownvalidation probability plus $10^{-8}$ and is therefore marked LOW_SUPPORT. A query is accepted on the fast route only when $q \geq 0.40, ~\pi_{\mathrm{F}} \geq \tau_{\mathrm{F}}$, and $\Delta \geq 0.02 ; ~q<0.40$ yields RETRY, and other cases proceed to AdaFace.

The final calibrator uses standardized fast/expert top scores and margins plus quality. It follows the same optimization procedure, with comparison operator $\geq$. The stored final threshold is 1.00000001 for every seed, so an expert-routed probe cannot satisfy the acceptance condition and the expert branch is rejection-only in this frozen evaluation. No gallery-dispersion or cross-model-agreement feature is used. No expert-assisted correction, degradation, or identity change occurred. The measured expert-fallback probability is

$p_{\mathrm{f}}=\frac{97}{1,675}=0.0579$

3.3 Permissioned audit record

The audit service recursively sorts associative payload keys before canonical JavaScript Object Notation (JSON) serialization. The data hash is computed as the SHA-256 digest of the canonical payload:

$h_{\text {data}}=S H A 256($canonical$($payload$))$     (4)

The block hash is computed from the length-prefixed block number, previous hash, data hash, and creation time. Length-prefix encoding prevents delimiter ambiguity; it does not authenticate the fields against an authorized or unauthorized database writer:

$h_{{block}}={SHA256}\left({enc}(b)\left\|{enc}\left(h_{{prev}}\right)\right\| {enc}\left(h_{{data}}\right) \| {enc}(t)\right)$        (5)

L0 is a conventional database log, L1 stores an independent record hash, L2 links sequential records, and L3 adds stored checkpoints. Full-chain validation recomputes the data and block hashes and verifies previous-hash linkage, continuity, schema version, and the verification-record link. These operations check the internal consistency of the presented records. SHA-256 is unkeyed: a writer able to modify records and recompute all affected subsequent hashes can construct a self-consistent replacement history without finding a hash collision. The archived evaluation establishes neither an independently authenticated chain head nor protection of such a reference outside the writable database. The source audit found checkpoint creation but no dedicated production checkpoint-validation command; stored checkpoints are not treated as authenticated trust anchors. Accordingly, the experiments support detection of the tested unrehashed inconsistencies only. They do not establish general tamper-evident security against coordinated rewriting, and checkpoint-validation and analytical break-even claims remain excluded.

5.1 Recognition workload characterization

Figure 2 visualizes the corresponding seed-level latency and Table 4 summarizes the recognition-workload statistics. QAC-OSFI reduced mean recognition-service latency by 49.26% relative to the reported AdaFace workload, but its P95 was similar and its P99 was higher. It used the fast route for 1,578/1,675 placements (94.21%) and the expert rejection route for 97/1,675 (5.79%). All 1,575 known placements were accepted through the fast route. Among 100 unknown placements, 97 were routed to the expert branch and rejected; three remained on the fast route and were falsely accepted.

2.png

Figure 2. Seed-level recognition latency; error bars denote standard deviations across five repeated seeds

Table 4. Recognition workload summary across repeated seeds (latency: Mean ± standard deviation)

The three placement-level false accepts were the same source image from one unknown subject repeated in seeds 101, 303, and 404. Thus, 1,672/1,675 correct placement decisions and 100% known rank-1 accuracy describe this dataset only. The evidence does not establish low-FPIR performance, independent-population accuracy, or external generalization.

5.2 Append-latency model and interaction sensitivity

Table 5 retains the numerical results of both specifications on the same reused validation subset. The append interaction variant had lower comparison-subset MAPE (40.07% versus 45.33%) and higher R² (0.629 versus 0.585). Because those results informed interaction selection, this apparent improvement is exploratory and cannot establish superior independent-test performance. Composite and validation-latency scores also remained poor on this subset. Neither specification has an untouched-test estimate in this study, and neither is presented as a deployment predictor.

Table 5. Exploratory model-comparison scores on the reused validation subset (19 configuration keys)

At the reference workload, the fitted interaction specification permits different concurrency slopes by ledger mechanism; its 25-worker/one-worker ratios ranged from 21.1 for L1 to 106.0 for L3. Figure 3 illustrates these interaction-aware fitted associations and their model-conditional confidence bands. The fitted-mean bands widen substantially at 25 workers. These model-conditional associations, and the 40.07% MAPE on the reused validation subset, do not account for selection on that subset and must not be treated as independently validated capacity-planning constants.

3.png

Figure 3. Interaction-aware fitted append latency; bands are nominal 95% fitted-mean intervals conditional on the model specification

5.3 Provenance audit and removal of unsupported optimization

The provenance audit identified three metrics that were not admissible for operational optimization. First, throughput_rps is algebraically identical to 10⁹ divided by composite latency for every warm row, so it is not an independently observed throughput window. Second, storage_bytes contains only 14 stepwise values and correlates with execution order (Spearman ρ = 0.996), indicating cumulative database state rather than configuration-specific storage growth. Third, the archived evidence does not verify that nominal chain_length equals the number of materialized linked records. The audit disposition is summarized in Table 6.

Table 6. Audit-based metric disposition

An archived configuration carrying nominal n = 10⁵ illustrates the final issue: its database-size snapshot was only 98,304 bytes and its 30-repeat validation mean was 1.95 ms. These values cannot support a full-100,000-block validation claim. Consequently, validation-based Pareto points, checkpoint break-even, storage optimization, and global-optimum statements are excluded. This study reports no operational Pareto set.

5.4 Integrity-control outcomes

Five isolated record/linkage inconsistencies were detected by the trusted verifier, and two structural actions were prevented by database constraints. Table 7 reports the mechanism and single-trial latency for transparency. No trial tested a coordinated history rewrite with consistent recomputation of subsequent hashes. These outcomes demonstrate only the specified consistency checks and structural preventions; they establish neither D = 1 as a general detection probability nor tamper-evident protection against an attacker with sufficient database write access.

Table 7. Single-trial consistency-check and database-prevention outcomes in the copied benchmark database

This study evaluated hash-chain-audited biometric verification within an explicit evidence-admissibility boundary. QAC-OSFI is specified as a MobileFaceNet fast route with quality-aware prototypes, validation-only logistic calibration, and an AdaFace rejection route. Its placement-level results are limited to 554 unique local images and 20 unknown identities. The ledger benchmark contributes 2,976 observations and supports descriptive append-contention analysis. On the reused 19-configuration validation subset, the interaction model yielded R² = 0.629 and MAPE = 40.07%; because this subset informed model selection, these are exploratory scores rather than unbiased independent-test estimates. A provenance audit excluded unsupported validation-latency, storage, throughput, checkpoint, and Pareto claims. Five unrehashed record/linkage inconsistencies were detected by a trusted verifier, and two structural attempts were prevented by database constraints, each in one trial. These outcomes do not establish tamper-evident protection against coordinated database rewriting and hash recomputation. Broader claims require untouched predictive evaluation after model freezing, independent biometric populations, verified measurement provenance, and repeated adversarial tests with explicitly protected trust references.

The anonymized performance measurements, model specifications, archived prediction files, integrity outcomes, environment metadata, and reproducibility documentation are available on Zenodo under DOI 10.5281/zenodo.21615345. The allocation matrix, interaction-sensitivity results, metric-admissibility audit, and tamper-outcome reclassification are supplied as supplementary material. Historical “holdout” labels in archived files denote the 19-key subset subsequently reused for model comparison and selection; they do not identify an untouched test set. Facial images, embeddings, names, identification numbers, and precise locations are not publicly released.