The Sensor's Own Flaws as Its Key: Pressure-Stable Fingerprints from Monolithic Membrane Arrays

Some passive wireless sensor tags put a pressure-sensing membrane and its readout electronics on the same silicon chip and run on a few microwatts harvested from radio waves. These tags have no way to prove who they are. They carry no stored key and no authentication, because a key memory or a cryptographic block costs more area, power, or fabrication steps than the chip can spare. The same tags also have a manufacturing nuisance. No two membranes come out identical, so each one has its own resting capacitance, and the factory has to measure and subtract that offset during calibration. This note proposes using that offset pattern as the tag's identity. Build the sensing element as a small array of nominally identical membranes. Read out the signs of the differences between the offsets of selected membrane pairs. Then keep only the pairs whose sign does not flip anywhere in the sensor's operating range of pressure and temperature. The resulting bits form a fingerprint of the physical transducer. The factory measurement that enrolls it is the same measurement that calibrates the sensor.

The claim is narrow and testable. We do not claim the membranes make a modeling-resistant "strong" physically unclonable function (PUF), meaning a circuit that answers an unlimited set of challenges unpredictably. We claim three things. First, a monolithic membrane array, read through the readout circuit the tag already has, yields a weak PUF: a fixed, die-unique bit string with uniqueness and reliability comparable to dedicated PUF circuits. Second, its added area and energy are smaller, because the transducer and the converter are reused. Third, it offers something a sealed PUF circuit cannot: if someone physically disturbs the membranes, by gluing, puncturing, or pre-stressing them, the key changes, so the sensor reveals that it has been tampered with.

Where this sits

Three groups of prior work set the stage. The first establishes the platform and the gap. A CMOS Pressure Sensor Tag Chip for Passive Wireless Applications builds a capacitive pressure membrane in 0.18 µm CMOS with a phase-locked-loop readout using 0.8 µW, at 3.2 µW total. It attributes chip-to-chip capacitance deviations to process variation and names the missing secure identification as an open problem. Design of an Embedded CMOS Temperature Sensor for Passive RFID Tag Chips reaches 92 nW but still needs a one-point calibration to remove process offsets. Both papers treat variation only as an impairment. We keep their readout architecture and turn the offset they subtract into the identity.

The second group supplies the methods. Process variation analysis for MEMS design models how geometric tolerances in micro-mechanical structures spread their stiffness and displacement, using Monte Carlo and ellipsoidal bounds. We reuse that kind of model to predict how much membrane-to-membrane mismatch to expect before committing to silicon. That paper bounds variation to suppress it; we want to measure it. A Reliable Physical Unclonable Function Based on Differential Charging Capacitors shows that capacitor mismatch read differentially gives a good PUF. Its capacitors are sealed MOS devices that cannot sense anything, and it was evaluated only in simulation. PUF Sensor: Exploiting PUF Unreliability for Secure Wireless Sensing contributes the enrollment idea we invert. It selects SRAM bits that predictably flip under heat in order to sense aging. We select membrane pairs that predictably do not flip under the measured quantity, so that sensing and identity can share one structure.

The third group constrains the design. PUF Modeling Attacks on Simulated and Silicon Data shows that linear-additive PUFs fall to logistic regression with roughly 650 challenge-response pairs. A differential capacitance between groups of membranes is exactly such a linear sum. This is why we restrict ourselves to a weak PUF whose bits never leave the chip. PUF-FSM: A Controlled Strong PUF warns that error-correction helper data leaks information. That warning motivates selecting stable pairs at enrollment rather than relying on heavy error correction.

Core Claim

Take $N$ membranes on one die, all exposed to the same pressure $P$ and temperature $T$. Model each membrane's capacitance as

$$C_i(P,T) \approx C^0_i + S_i P + \alpha_i (T - T_0),$$

where $C^0_i$ is the membrane's resting offset, $S_i$ its pressure sensitivity, $\alpha_i$ its temperature coefficient, and $T_0$ the enrollment temperature. The sign of the difference $C_i - C_j$ is a stable key bit if the gap between offsets dominates the gaps in sensitivity and temperature drift over the whole operating range:

$$|C^0_i - C^0_j| > |S_i - S_j|\, P_{\max} + |\alpha_i - \alpha_j|\, \Delta T_{\max} + k\sigma_n,$$

where $P_{\max}$ and $\Delta T_{\max}$ are the operating extremes, $\sigma_n$ is the readout noise, and $k$ is a safety factor.

Enrollment measures $C^0_i$ and $S_i$ at two pressures and two temperatures. That is the same data a calibration table needs. From it we select the disjoint pairs that satisfy the inequality. Because a common pressure moves every membrane in the same direction, the differential readout cancels most of the signal the sensor is built to measure. Only the sensitivity mismatch remains, and the selection rule guards against it. This resolves the tension between a sensor that must respond strongly to its environment and a PUF that must not. The deployed tag senses with the array average and derives its key from the residual pattern.

Research Program

The work proceeds in three stages. Stage 2 is gated on the outcome of stage 1.

The first stage is a coupled simulation. It combines geometric and residual-stress variation models for the membranes, in the style of Process variation analysis for MEMS design, with a transistor-level model of the readout. This answers the go/no-go question the file cannot: is the within-die spread of $C^0_i$ large compared with the readout's resolution? The Deng readout resolves 7.4 bits over a 0.72 pF span, which works out to about 5.6 fF per step. If the within-die offset spread is only a step or two, the idea needs longer integration or fails.

The second stage is fabrication. We would tape out test dies in a 0.18 µm CMOS process with post-CMOS membrane release, following Deng et al. Each die carries an array of 16 to 64 small membranes, a switch matrix that routes any pair into the existing differential readout, and an equal number of unreleased "sealed" membranes with identical electrodes as a control. The third stage is a characterization campaign over many dies, across pressure, temperature, harvested RF power, and deliberate tampering.

The authentication protocol on top of the key, such as a lightweight message authentication code, is shared by every baseline. We therefore account its cost once and compare only the cost of generating the key.

Baselines

The naive reference point is the existing tag with no identity. The next step up is the same array keyed from all pairs without stability selection. That shows how often unselected bits flip once pressure varies, which is the error rate a fuzzy extractor would otherwise have to absorb.

The strongest dedicated competitors are co-located PUF circuits placed beside an unchanged sensor.

The first is the differential charging-capacitor PUF of A Reliable Physical Unclonable Function Based on Differential Charging Capacitors. It is the closest mechanism, and it reports 50.03% uniqueness, 97.70% reliability over 0–80 °C, and 750 fJ/bit at 20.86 µW, in simulation only.

The second is the hybrid ring-oscillator PUF of A Low-Power Hybrid RO PUF With Improved Thermal Stability for Lightweight Applications. It has the best reported silicon reliability: 99.84% average and 97.28% worst case over −40 to 120 °C, 50.42% inter-die distance, 250 µm², and 32.3 µW per challenge-response pair.

The third is TV-PUF : A Fast Lightweight Aging-Resistant Threshold Voltage PUF. It has the lowest power, 0.18 µW per bit at 586 transistors per bit.

For a fair comparison we would lay out the capacitor PUF and the ring-oscillator PUF on the same test die and process. Each would be sized to deliver the same number of post-selection key bits. All three would be measured with the same instruments over the same temperature and supply sweeps, and given the same enrollment budget of four measurement conditions. TV-PUF numbers come from simulation at other nodes, so we treat it only as a reference point for area and power.

Our design beats the dedicated PUFs if two conditions hold together. Its worst-case reliability must match or exceed 97.28% over the sensor's full pressure range as well as temperature. Its incremental area and energy per key bit must also be lower.

For the calibration side, there is no reported result found on Lacuna for calibrated pressure accuracy on the Deng pressure tag. Our reference is therefore that tag's architecture calibrated in the conventional way on our own dies. The temperature tag's −0.7/0.6 °C after one-point calibration indicates the scale of accuracy a one-shot enrollment must preserve.

Two ablations isolate the key ingredients. The first comes from the sealed-membrane control. If the sealed electrodes give the same uniqueness and stability as the released membranes, then the entropy comes from ordinary electrode geometry, not from the transducer. In that case the tamper-evidence claim collapses, and the design is just a capacitor PUF in disguise. The second ablation removes the stability-selection rule while keeping everything else fixed.

Why not just…?

Why not just place a small dedicated PUF next to the unchanged sensor? This is the real threat, and we do not argue around it. The capacitor and ring-oscillator PUFs above are mature, and 250 µm² is small next to a pressure membrane. Our proposal wins only if two experiments both come out in its favor. First, the membrane key must cost measurably less area, energy, and test time, because enrollment is folded into calibration. Second, it must show tamper evidence that a sealed PUF lacks. If the tampering experiments show the key surviving membrane disturbance, or if the switch matrix eats the savings, the side-by-side PUF is the better engineering choice. We would report that outcome as such.

Why not just store a factory key in embedded non-volatile memory? The standard-CMOS-plus-release process used by Design of a Humidity Sensor Tag for Passive Wireless Applications and the pressure tag has no flash layers. Adding them means extra masks, and a stored key can be read by probing. This alternative mainly defines the cost the PUF approaches are trying to avoid, so we treat it as a cost reference rather than run it.

Why not just send the raw offset vector and let the verifier match it, as the SRAM PUF sensor does? PUF Sensor: Exploiting PUF Unreliability for Secure Wireless Sensing moves all statistics off-chip, which is the cheapest design. But an eavesdropper who records two readings can recover every $C^0_i$ and $S_i$ and replay plausible data. For the same reason, challenge-selected group differences would be learnable, by the linear-model results of PUF Modeling Attacks on Simulated and Silicon Data. We will show this concretely by fitting a logistic-regression attacker to transmitted pair readings. That experiment justifies keeping the key bits on chip.

Why not just use a fuzzy extractor instead of selecting stable pairs? Error correction would tolerate flipping bits, but it costs digital area the tag lacks, and its helper data leaks, as PUF-FSM: A Controlled Strong PUF shows. The unselected-pairs ablation measures how much error correction would be needed. If that residual error is small enough for a tiny repetition code, then pair selection and light correction should be combined rather than treated as alternatives.

Proposed Experiments

Four checks are decisive. The first is the stage-1 measurement of within-die membrane offset spread against readout noise. Everything else depends on its ratio. The second sweeps pressure from 0 to 600 kPa and temperature from −40 to 120 °C on 30 or more dies. It reports, against the three baselines measured on the same dies, the post-selection bit yield per membrane, intra-die and inter-die Hamming distance, worst-case reliability, and NIST SP 800-22 randomness on the pooled bits. The third checks whether a single enrollment gives calibrated pressure accuracy matching conventional calibration of the same chips. The fourth tampers with the membranes, applying adhesive, local loading, and puncture to some of them, and measures how often the key changes. We count detection only when the change is not also triggered by legitimate pressure.

The decisive outcome is any of the following: a key stable across the sensing range at under a dedicated PUF's incremental cost, with tamper detection; an entropy deficit; or a demonstration that the sealed control does just as well.