Cross-Layer Roots of Trust: Integrating Biometrics, PUFs, and Hardware Obfuscation

Nima Karimian

Source-linked AI summary

Cross-Layer Roots of Trust: Integrating Biometrics, PUFs, and Hardware Obfuscation

Nima Karimian

arXiv:2608.21643v1Aug 21, 2026cs.CR

TL;DR

The paper addresses the limited systematic understanding of how human, device, and hardware-function trust compose in embedded and edge systems. It develops a unified framework covering complete primitive pipelines and their pairwise and three-way compositions, concluding that cross-layer reliability, leakage, and attack chains require end-to-end evaluation.

Problem
Trust in the requesting human, participating physical device, and authorized hardware function is usually studied separately, although their interfaces can create coupled failures and attacks.
Method
The survey decomposes biometrics, PUFs, and hardware obfuscation into complete processing chains and systematically analyzes their assumptions, metrics, failure modes, and pairwise and three-way compositions.
Results
The survey establishes a unified human–device–function trust architecture in which authorization depends jointly on human trust, device trust, and protected-function authorization.
Takeaways & Limitations
End-to-end assessment must include cross-layer reliability, key derivation, helper-data handling, modeling and side-channel attacks, functional activation, and authorization.
Takeaways & Limitations
Fixed quantization offers low storage and computation but can involve a security trade-off between reconstruction reliability, helper-data leakage, key length, and entropy.

Abstract

from arXiv · show

Modern cyber--physical, Internet-of-Things (IoT), wearable, and edge systems increasingly require trust in three distinct entities: the human requesting access, the physical device executing the computation, and the hardware function that is permitted to operate. These requirements are usually studied in separate communities. Biometrics establish human identity but remain vulnerable to presentation attacks, intra-user variability, template leakage, and limited revocability. Physical unclonable functions (PUFs) provide device-specific physical identity and on-demand secret derivation, yet must address environmental instability, helper-data exposure, side channels, and modeling attacks. Hardware obfuscation and logic locking condition correct circuit behavior on an activation secret, but face oracle-guided, approximate, structural, removal, and physical attacks. This survey develops a unified human--device--function view of trust. We first decompose each primitive into its complete processing chain and identify the corresponding security assumptions, implementation mechanisms, and evaluation metrics. We then formalize pairwise compositions---biometric--PUF, PUF--obfuscation, and biometric--obfuscation---and a three-way architecture in which correct functionality is bound jointly to an authorized user and a genuine device. Particular attention is given to biometric key reconstruction, PUF stabilization and modeling resistance, logic-locking attack evaluation, cross-layer error propagation, enrollment trust, key lifecycle, and interface leakage. The survey concludes with a taxonomy and research agenda for revocable human--device credentials, compositional security, leakage-aware integration, reconfigurable activation, and standardized end-to-end evaluation.

I. INTRODUCTION

The survey frames authorization as a human–device–function trust problem because biometrics, PUFs, and hardware locking provide distinct but interacting trust conditions. It organizes these primitives as composed processing pipelines, emphasizing cross-layer attacks, error propagation, and end-to-end evaluation.

I. INTRODUCTION: Biometrics, PUFs, and logic locking establish trust respectively in the human, physical device, and authorized hardware function.The architecture treats these as complementary roots rather than a single credential.
I. INTRODUCTION: Security does not compose automatically: legitimate users may access counterfeit devices, genuine devices may serve unauthorized users, and exposed keys may bypass locked circuits.Connected interfaces become part of the attack surface.
I. INTRODUCTION: The survey models authorization as A = T_H ∧ T_D ∧ T_F, without assuming statistical independence among human, device, and function trust.Failures can propagate through key derivation, device binding, and hardware activation.
I. INTRODUCTION: Each primitive is examined through its complete processing pipeline, implementation assumptions, security properties, failure modes, evaluation metrics, and pairwise or three-way compositions.The framework targets systems where an authorized user interacts with an authentic device to activate a protected function.
I. INTRODUCTION: Biometric authorization spans sensing, quality control and PAD, preprocessing, feature extraction, protection, matching or recovery, and final decision.Every stage introduces distinct security exposure, cost, power, latency, and scalability constraints.
I. INTRODUCTION: Biometric key generation requires exact binary-key recovery, making intra-user variability and reliability protection more consequential than in similarity-based verification.Its costs can include quantization, reliability estimation, secure-sketch or fuzzy-extractor operations, decoding, hashing, and protected storage.

6) Matching or Key Recovery:

Matching compares probe and enrolled biometric representations against a threshold, whereas key recovery reconstructs enrollment material from noisy measurements and helper data. The complete stage remains security-critical because attacks, implementation costs, and upstream errors can affect authorization.

Verification: Conventional verification accepts a probe when its similarity score crosses threshold τ, trading false acceptance against false rejection.FAR measures impostor acceptance, while FRR measures legitimate-user rejection.
Key Recovery: Biometric key-generation systems instead reconstruct the enrollment key from a noisy binary sequence using helper data and a recovery operation.The recovered key can be verified by comparing its protected representation with the enrolled value.
Cost: Matching scales with database size, while key recovery avoids template search but adds error correction, helper-data processing, hashing, or cryptographic verification.Direct comparison of N m-dimensional templates requires approximately Cmatch = O(Nm).
Authorization: Authorization is the final security-critical stage: decision overrides, threshold manipulation, privilege escalation, or leakage can nullify preceding protections.In the integrated architecture, biometric errors or leakage can propagate into PUF challenge generation and hardware authorization.
Key Recovery: Key-generation evaluation must jointly report reliability, entropy, key length, inter-user separation, helper-data overhead, and implementation costs.These metrics capture reproducibility, unpredictability, separation, and resource demands across the recovery pipeline.
Verification: Threshold selection creates a security–usability trade-off: increasing τ generally lowers FAR but raises FRR.Deployment studies should report FAR and FRR at application-specific thresholds, especially when hardware activation requires very low FAR.

IV. BIOMETRIC KEY GENERATION THROUGH QUANTIZATION

Quantization converts extracted biometric features into binary key material, using fixed or adaptive mappings. Fixed quantization is simple and constant-length but is sensitive to measurements near global decision boundaries.

Overview: Biometric key generation must repeatedly map same-user measurements to identical binary material or use error correction to reconstruct the enrollment key.This exact-recovery requirement is stricter than threshold-based biometric verification.
Overview: Quantization applies feature extraction followed by a mapping from continuous feature values to binary symbols.The pipeline represents features as z_u = f(x_u) and keys as K_u = Q(z_u).
Fixed and Adaptive Quantization: Fixed quantization uses globally shared intervals, whereas adaptive quantization uses enrollment statistics, guard bands, or feature selection to exclude unstable regions.Fixed boundaries require little user-specific metadata; adaptive boundaries or selection criteria are user-dependent.
Fixed Quantization: A q-bit fixed quantizer partitions each feature range into M = 2^q intervals and maps the containing region to a q-bit word.For feature j, region r is selected when β_j,r ≤ z_u,j < β_j,r+1.
Enrollment and Protection: Fixed-quantization enrollment applies the same boundaries to representative enrollment features, concatenating symbols into a constant-length key L_K = mq.Protected verification values and optional helper data can replace direct storage of the raw key.
Authentication: During authentication, the same feature extraction and global quantization pipeline reconstructs a key, with helper data supporting recovery when residual errors remain.No guard-band test is normally applied, reducing runtime complexity but increasing sensitivity to intra-user variation.

3) Reliability Under Fixed Quantization:

Fixed quantization offers simple, low-storage implementation but makes reliability depend on each user’s distance from global boundaries and on the population distribution of generated symbols. Adaptive quantization addresses this through reliability filtering and guard bands, while trading away unstable or low-entropy features.

Reliability: Fixed-boundary reliability varies with a user’s mean location: features near interval centers are more stable than features near boundaries.Users with the same feature variance can therefore experience different reconstruction reliability.
Reliability: A small per-bit error probability can make exact reconstruction of a long key unlikely, approximately KR ≈ (1 − p_b)^L_K.The approximation assumes independent bit errors without error correction.
Entropy and Separation: Uniform intervals do not guarantee uniform symbols; nonuniform biometric distributions can reduce effective key entropy despite nominally contributing q bits per feature.Inter-user Hamming distance should be interpreted jointly with entropy because distance near 0.5 does not prove independence or prediction resistance.
Implementation: Fixed quantization costs approximately O(mM) comparisons with linear search or O(m log M) with binary lookup, and its boundary storage is independent of enrolled-user count.User-specific storage can be limited to protected verification data and optional helper data.
Advantages: Fixed quantization provides deterministic mapping, constant key length, small user-specific storage, and straightforward hardware realization.The same quantizer can be implemented for every user.
Limitations: Fixed quantization lacks user-specific reliability adaptation, so small biometric variations can cause bit transitions even when conventional matching still identifies the same user.Robustness is consequently shifted toward preprocessing, fuzzy extraction, and error correction.
Adaptive Quantization: Adaptive quantization retains features only when reliability and population-level entropy exceed thresholds, expressing a reliability–entropy trade-off.Highly stable but nondiscriminative features and distinctive but unstable features are both excluded.
Adaptive Quantization: Guard bands reject measurements near quantization boundaries instead of forcing them into neighboring codes.A feature contributes a bit only when its minimum boundary margin meets the required safety margin.

1) Adaptive Guard Bands:

Adaptive quantization builds a user-dependent map that retains reliable, discriminative feature regions and excludes measurements near unstable boundaries. Enrollment estimates feature statistics, selects suitable intervals, and prepares parameters and helper data for later key reconstruction.

Adaptive Guard Bands: Adaptive quantization estimates each user’s feature behavior during enrollment and retains features meeting stability and discriminability requirements.The resulting map is user dependent rather than governed by shared global boundaries.
Adaptive Guard Bands: Enrollment collects repeated measurements to estimate feature means and variances, with larger variance indicating greater risk of boundary crossing.These statistics support user-specific reliability margins and interval selection.
Adaptive Guard Bands: Guard bands exclude measurements near decision boundaries instead of forcing uncertain observations into neighboring binary states.This treats boundary-adjacent observations as erasures during authentication.
Adaptive Guard Bands: Feature selection balances reliability against population entropy, rejecting unstable high-entropy features and predictable low-entropy features.Retained features are partitioned into selected intervals mapped to binary codewords.
Adaptive Guard Bands: The raw key concatenates codewords from retained features, while its length may vary by user according to the number of reliable quantization states.Enrollment also produces quantization parameters, helper information, and a protected verification representation rather than storing the raw key.
Adaptive Guard Bands: Authentication reprocesses selected features with stored parameters, reconstructs a noisy sequence, applies error correction when needed, and verifies the protected representation.Only enrollment-selected features are evaluated, using the stored user-specific quantization parameters.

C. Fixed Versus Adaptive Quantization: RQ-Driven Analysis

Fixed quantization favors simplicity and global parameters, whereas adaptive quantization trades greater enrollment complexity for improved robustness under variability and noise. Evaluation must jointly consider reliability, entropy, retained key length, discriminability, operating conditions, and implementation cost.

C. Fixed Versus Adaptive Quantization: RQ-Driven Analysis: Fixed quantization assigns every observation to predefined intervals, so small boundary perturbations can change the generated bit.Adaptive quantization instead discards boundary-adjacent measurements through guard bands.
C. Fixed Versus Adaptive Quantization: RQ-Driven Analysis: Adaptive quantization should be judged by whether reduced intra-user BER justifies the loss of usable features and key length.The principal reliability metric is RelK = 1 − BERintra, paired with reporting of retained length.
C. Fixed Versus Adaptive Quantization: RQ-Driven Analysis: Noise-aware ECG results reported high reliability under severe synthesized noise for NA-IOMBA, while conventional IOMBA degraded more strongly.This comparison is presented as evidence for evaluating robustness under controlled noise conditions.
C. Fixed Versus Adaptive Quantization: RQ-Driven Analysis: Adaptive quantization adds enrollment costs for statistics, selection, interval optimization, and optional error-correction design.Online processing can nevertheless be lower because only the selected features are processed.
C. Fixed Versus Adaptive Quantization: RQ-Driven Analysis: Strategy choice depends on application priorities spanning implementation simplicity, enrollment cost, runtime processing, reliability, environmental noise, and error-correction burden.For hardware-rooted systems, unreliable biometric bits can propagate into PUF challenges, cryptographic keys, or obfuscation keys.

D. Biometric Challenges and Limitations

Biometric outputs become stricter security inputs when they unlock hardware, requiring exact key reconstruction and protection against variability, spoofing, leakage, and limited revocability. PUFs similarly transform device-specific physical variation into stabilized, corrected keys or challenge–response credentials, with reliability–cost and helper-data security trade-offs.

D. Biometric Challenges and Limitations: Biometric hardware unlocking amplifies risks from intra-user variation, poor revocability, presentation attacks, side channels, and environment-correlated errors.These factors can invalidate simplistic independence assumptions in joint-system analysis.
D. Biometric Challenges and Limitations: A PUF derives device identity from manufacturing variation rather than a conventionally stored secret, producing challenge–response behavior tied to the physical instance.The response also depends on challenge, environment, and aging effects.
D. Biometric Challenges and Limitations: Practical PUF processing digitizes physical behavior, filters unstable responses, generates helper data, performs error correction, and derives a stable key or credential.The complete path is summarized from the physical entropy source through raw and stable responses to final output.
D. Biometric Challenges and Limitations: PUF stabilization through repeated measurement or majority voting improves reliability but increases enrollment time, latency, energy, and metadata storage.Error correction adds redundancy, decoding logic, helper data, and latency as the correction radius grows.
D. Biometric Challenges and Limitations: Helper-data manipulation can alter decoding behavior or reduce effective security, so passive exposure alone is insufficient for PUF fuzzy-extractor analysis.Explicit helper-data masking has been proposed in response to this threat.

B. Weak and Strong PUFs

Weak PUFs expose limited response spaces suited to local key generation, while strong PUFs expose large challenge spaces for authentication. PUF evaluation must combine statistical quality, environmental stability, and explicit resistance to modeling and other attacks.

B. Weak and Strong PUFs: Weak PUFs expose limited effectively independent responses and are mainly suited to local key generation or device fingerprinting.SRAM and many memory-based PUFs commonly fall into this category.
B. Weak and Strong PUFs: Strong PUFs provide much larger challenge spaces and are intended to support authentication through many challenge–response pairs.Arbiter-derived constructions are canonical examples.
B. Weak and Strong PUFs: A large challenge–response space does not guarantee unpredictability because machine-learning models can approximate several strong PUF constructions from observed CRPs.Neural-network attacks extended modeling capability against multiple arbiter-PUF variants.
B. Weak and Strong PUFs: PUF assessment must report statistical response quality alongside environmental stability, helper-data security, side-channel exposure, and modeling resistance.ISO/IEC 20897 standards provide baseline requirements and evaluation methods, but application-specific attacks require additional testing.
B. Weak and Strong PUFs: Uniqueness, reliability, uniformity, and bit aliasing characterize device separation and response bias, but none alone establishes cryptographic independence or entropy.Ideal balanced-response references include Uniqideal = 0.5 and Unifideal = 0.5.
B. Weak and Strong PUFs: Reliability should be measured across operating conditions and aging, while modeling experiments should report both prediction accuracy and CRP sample complexity.Reporting only attack accuracy can obscure how many exposed CRPs are needed to reach it.

E. Helper-Data and ECC Overhead

PUF key-generation cost includes stable-bit selection, helper-data storage, error-correction redundancy, measurement, and implementation overhead. Reliability improvements can increase latency, energy, storage, and hardware cost.

Storage overhead: PUF key-generation storage includes selected-response data, ECC redundancy, and helper data, while the derived secret need not be stored persistently.The relevant quantities include Lraw, Lsel, ECC parameters, and Shelper = |W|.
Cross-layer correction: Combining biometric and PUF layers can require separate error-correction procedures, motivating joint reliability selection before expensive decoding.The paper identifies total correction latency as a cross-layer cost and recommends screening reliability jointly.
Platform dependence: PUF evaluation cost depends on whether the physical structure already exists in the target platform and on the required measurement process.Memory PUFs can reuse existing memory, whereas DRAM PUFs may require repeated accesses and controlled retention or timing conditions.
Latency: Increasing the RO-PUF measurement window can improve frequency-estimation confidence but directly increases authentication latency.The latency model includes stimulation, measurement, digitization, filtering, ECC, and extraction stages.
Security boundaries: PUF uniqueness and reliability do not guarantee cryptographic unpredictability, especially for strong PUFs vulnerable to response-modeling attacks.Environmental variation, helper-data manipulation, side channels, and fault injection also require explicit threat-model coverage.

3) Locked Circuit and Key-Dependent Behavior:

A locked circuit restores intended behavior only with the authorized activation key, while incorrect keys should corrupt functionality across a meaningful input region. Evaluation must therefore cover exact and approximate attacks, structural removal, physical key exposure, and implementation cost.

Key-dependent behavior: The authorized key satisfies FL(x, K*) = F(x), whereas incorrect keys should differ from F(x) over a substantial portion of the input space.Correct activation is necessary but insufficient unless wrong-key behavior is meaningfully corrupted.
Oracle-guided evaluation: SAT attacks use distinguishing input patterns and oracle outputs to eliminate inconsistent key candidates until remaining keys are functionally equivalent.Recommended reporting includes TSAT, NDIP, and Nquery rather than key length alone.
Approximate attacks: Approximate attacks can recover keys that are not exact but reproduce most intended functionality, particularly when corruption is confined to rare input patterns.AppSAT exemplifies early termination before exact SAT convergence.
Structural attacks: Structural attacks can identify and remove protection logic without recovering the intended key, so functional corruption must be paired with removal-success analysis.Topology, gate types, fan-in/fan-out, symmetry, and synthesis artifacts may expose locking structures.
Runtime activation: Activation security depends on protecting the key across storage, loading, transfer, and hardware interfaces, including when biometric or PUF-derived keys are used.Exposure between the key-derivation unit and locked hardware can negate individually secure upstream primitives.

2) Key Sensitivity:

Logic-locking security is a multi-objective property shaped by key influence, attack resistance, wrong-key corruption, implementation overhead, synthesis effects, and activation-path protection. No single metric or locking family captures this behavior reliably.

Key sensitivity: Low key sensitivity indicates that some activation bits have little functional influence and may be structurally or functionally weak.Sensitivity can be examined by inverting individual bits of the authorized key.
Evaluation metrics: A rigorous evaluation should report exact and approximate attack success, query complexity, distinguishing inputs, timeouts, key sensitivity, corruption, and implementation overhead.The relevant attack metrics include Tattack, Nquery, NDIP, PKR, and Papprox.
Implementation overhead: Hardware protection must be assessed against unlocked-design area, delay, and power baselines, while runtime cost includes key loading or reconstruction.Reconfiguration time can be non-negligible for FPGA, eFPGA, and bitstream-based protection.
Integrated architecture: Functional authorization cost cannot always be separated from the human and device roots of trust in the integrated architecture.The architecture treats authorization as a coupled cost across the complete chain.
Design trade-offs: Logic-locking families trade solver resistance, wrong-key corruption, structural protection, and implementation complexity rather than optimizing one universally dominant property.Random XOR/XNOR, SARLock- or Anti-SAT-like, and approximation-aware designs occupy different points in this trade-off space.
Implementation effects: Security can be overestimated when evaluation stops at RTL because synthesis and physical implementation may alter protection topology and key observability.Post-synthesis netlists, and ideally post-layout implementations, should be examined.

1) Reliability Propagation:

Composed human–device–function systems must stabilize biometric and PUF information before cryptographic fusion because errors and leakage propagate across layers. The resulting activation path binds functionality to reconstructed human and device credentials but introduces coupled reliability and integrity requirements.

Error propagation: Direct XOR fusion propagates biometric and PUF bit errors, while a cryptographic KDF can produce unrelated output after even one incorrect input bit.The cited error equation applies to direct bitwise fusion, not to cryptographic KDF error probabilities.
Stable reconstruction: Both biometric and PUF sources should be reconstructed into stable forms before cryptographic fusion.Successful binding depends on successful reconstruction at both layers, with direct measurement required when errors are correlated.
Enrollment and runtime: Enrollment establishes biometric reliability parameters and PUF response reliability information, while raw biometric measurements and raw device secrets need not be stored as plaintext.Runtime exposes the final binding secret only after both factors are successfully reconstructed.
PUF–obfuscation binding: PUF-derived activation keys bind correct hardware functionality to the enrolled physical device, leaving a copied netlist locked on another device.The transferred device is expected to generate a different PUF response and consequently a different activation key.
Interface protection: The KDF-to-locking interface must remain inside the trusted hardware boundary, and helper-data integrity is required because altered helper data can corrupt activation-key reconstruction.Probing, scan access, debug interfaces, side channels, or malicious helper-data modification can bypass or disrupt the binding.
Biometric–obfuscation binding: Biometric-controlled hardware activation requires exact biometric reconstruction because approximate similarity is insufficient for deriving the correct functional key.Quantization, fuzzy extraction, secure sketches, or other error-tolerant reconstruction must precede activation.

D. Biometric + PUF + Obfuscation: C7 Human–Device– Function Binding

C7 binds correct functionality to an authorized human and intended physical device through biometric reconstruction, PUF-derived device material, and protected hardware activation. The architecture separates stable activation from session freshness and evaluates successful activation and unauthorized activation across the full chain.

C7 requires correct functionality only when an authorized human interacts with the intended physical device and the joint credential activates the protected function.
The runtime chain derives biometric material, uses it to define a PUF challenge, fuses stable biometric and PUF components, and derives a domain-separated activation key.
A stable human–device binding secret supports static locking, while fresh nonces generate session authenticators without changing the static activation key.
A wrapped-key variant keeps the logic-locking key fixed and allows reconstructed binding material to decrypt it for circuit activation.
End-to-end activation requires successful biometric, PUF, and activation-key reconstruction, with the design targeting both LAFR →0 and UAR →0.

9) Implementation Trust Boundaries:

The integrated architecture treats cross-layer interfaces as explicit trust boundaries because leakage, modeling, and compromise can propagate across biometric, PUF, KDF, and hardware layers. It therefore calls for composition-aware evaluation, lifecycle protection, and standardized end-to-end benchmarks.

Implementation Trust Boundaries: Each interface carries security-sensitive material, and exposure of the KDF-to-locking value can bypass both biometric and PUF roots of trust.
Implementation Trust Boundaries: Short-lived intermediate values and protection against software buffers, debug interfaces, scan chains, and configuration buses reduce exposure during activation.
Cross-Layer Failure Propagation: Compromise can form attack chains from biometric leakage to PUF querying and modeling, derived binding secrets, and hardware activation-key recovery.
Cross-Layer Failure Propagation: A secure composition requires formal models and benchmarks because individual biometric, PUF, and locking security claims do not automatically compose.
Lifecycle and Revocation: Ownership transfer and device replacement must support new human–device bindings without exposing reusable biometric information or replacing physiological traits.
Joint Optimization: Joint optimization can select biometric and PUF features using entropy, reconstruction error, and leakage penalties rather than optimizing both layers independently.
Standardized Benchmarks: PUF evaluation should vary attacker knowledge, CRP budget, architecture knowledge, environmental side information, and adaptive querying.
Standardized Benchmarks: A standardized benchmark should combine biometric, bio-key, PUF, locking, PPA, side-channel, enrollment, and revocation metrics for the same end-to-end implementation.