A Physical Unclonable Function (PUF) is a silicon fingerprint: a circuit whose microscopic manufacturing variations decide, for each input bit-string (a "challenge"), a race between two signal paths, producing an output bit (a "response") that is unique to that chip and hard to predict in advance. The catch, demonstrated over a decade of attacks, is that although the physics is unclonable, the input-output mapping often is not: an eavesdropper who collects enough challenge-response pairs can train a software clone that answers unseen challenges just like the hardware (Modeling attacks on physical unclonable functions). The problem is that on the hardened designs that actually get deployed — chains whose responses are XORed together, loops that feed intermediate signals forward, bits "interposed" from an upper circuit into a lower one's challenge — every successful attack today is either a heuristic decomposition (Splitting the Interpose PUF: A Novel Modeling Attack Strategy) or a derivative-free search (Modeling Strong Physically Unclonable Functions with Metaheuristics), and these cost hundreds of millions of intercepted responses and days of compute.
I propose to attack these designs with machinery built for the exact same mathematical obstacle in a different field. A PUF's arbiter is a hard threshold: the response is a step function of the internal delay difference. Spiking neural networks face the identical structure — a neuron emits a binary spike when its membrane potential crosses a threshold — and the field's standard remedy is surrogate gradient learning: run the forward pass with the true discrete step function, but replace its nonexistent derivative in the backward pass with the derivative of a smooth stand-in (Surrogate Gradient Learning in Spiking Neural Networks; Spatio-Temporal Backpropagation for Training High-performance Spiking Neural Networks). No one has pointed this machinery at hardware security primitives. The bet: a structured differentiable clone of a strong PUF, trained end-to-end through its hard thresholds with surrogate gradients, will need orders of magnitude fewer challenge-response pairs than the incumbent attacks, because gradients — not bit-guessing heuristics or evolutionary sampling — carry information about the internal, unobserved threshold decisions.
Core Claim
The claim is that the internal hard decisions inside strong PUFs, which force today's attacks into divide-and-conquer guessing or derivative-free evolution, are not an obstacle but exactly the situation surrogate gradients were invented for. Concretely, I will build a "threshold-faithful clone": one learned linear delay model per arbiter chain, executed in the forward pass with the true step function, and composed through the design's XOR, feed-forward, and interpose wiring as a differentiable computational graph whose only non-differentiable operators are handled by surrogate derivatives in the backward pass. For a single chain with challenge $C$, transformed feature vector $\boldsymbol{\Psi}(C)$, and learned stage-delay vector $\mathbf{w}$:
$$\text{forward: } \hat{r} = H\!\left(\mathbf{w}^\top \boldsymbol{\Psi}(C)\right), \qquad \text{backward: } \frac{\partial \hat{r}}{\partial \mathbf{w}} \;\to\; \sigma'\!\left(\mathbf{w}^\top \boldsymbol{\Psi}(C)\right)\, \boldsymbol{\Psi}(C)$$
where $H$ is the Heaviside step and $\sigma$ a smooth surrogate (sigmoid or piecewise-linear, per the robustness findings in Spatio-Temporal Backpropagation for Training High-performance Spiking Neural Networks). The decisive difference from prior attacks is compositional. Logistic regression already is a smooth relaxation of one threshold — that is why it breaks simple Arbiter PUFs — but it cannot propagate error through a threshold, which is why the Interpose attack must randomly guess the intermediate interposed bit and iterate (Splitting the Interpose PUF: A Novel Modeling Attack Strategy), and why the chosen-challenge attack on noise-free XOR designs must fall back to CMA-ES correlating a non-flipping heuristic (Breaking XOR Arbiter PUFs without Reliability Information). Surrogate gradients let one training signal flow from the observed XOR or interposed output all the way back to every chain's parameters, jointly, in one pass.
Two known failure modes of surrogate training shape the design rather than arguing it away. The forward-backward mismatch accumulates in deep or recurrent compositions (Hybrid Macro/Micro Level Backpropagation for Training Deep Spiking Neural Networks.), and hand-fixed surrogate curves zero out informative gradients in sub-threshold regions — SURGE: Surrogate Gradient Adaptation in Binary Neural Networks fixes this in 1-bit quantized networks with a detached full-precision auxiliary path and an adaptive gradient scaler, and Differentiable hierarchical and surrogate gradient search for spiking neural networks fixes it with a learned mixture of surrogate functions (DGS). I will import both: an auxiliary continuous path per chain à la SURGE, and DGS-style adaptive surrogate parameters, treating the PUF clone as the first hardware-security testbed for adaptive surrogate learning.
Research Program
The work has three parts. First, build the differentiable PUF-clone library: parameterized forward models for the Arbiter, XOR Arbiter, Feed-Forward Arbiter, and Interpose architectures matching the additive linear delay simulation used across the attack literature (Modeling attacks on physical unclonable functions; Modeling Strong Physically Unclonable Functions with Metaheuristics), each exposing its thresholds as surrogate-differentiable operators, implemented in PyTorch on top of the simulation patterns available in BindsNET. Second, run the attack study: for each architecture family, train the clone on an increasing CRP budget and record accuracy, convergence time, and total CRPs to target accuracy, against the incumbents under matched budgets. Third, analyze why it works or fails: measure the forward-backward mismatch empirically (agreement between surrogate-predicted and true output flips near thresholds) and test whether the known SNN pathology (Hybrid Macro/Micro Level Backpropagation for Training Deep Spiking Neural Networks.) or the known SNN cure (Differentiable hierarchical and surrogate gradient search for spiking neural networks) dominates in this regime — this directly addresses the open tension of Gap 5, namely whether gradient approximation error or gradient guidance wins on brittle cryptanalysis losses.
A secondary use of the same machinery targets the hardest defended target in the file: the trinary quadruple response PUF with challenge obfuscation, whose authors project CMA-ES would need approximately 1875 years to reach 60% accuracy (Reliable and Modeling Attack Resistant Authentication of Arbiter PUF in FPGA Implementation With Trinary Quadruple Response). Its metastability outputs (0, 1, X) are a three-way thresholding, which a pair of surrogate-differentiable thresholds models naturally; and proximity to the metastable boundary is already known to leak delay magnitude (Breaking XOR Arbiter PUFs without Reliability Information), giving the clone a gradient-compatible supervision signal. I treat this as stretch goal, not core claim.
Proposed Experiments
All core experiments use the synthetic CRP generators standard in the field, since the literature notes no open PUF datasets exist in the catalog and studies overwhelmingly generate synthetic CRPs via additive linear models; hardware validation on a Xilinx Nexys 4 DDR board follows the protocol of Breaking XOR Arbiter PUFs With Chosen Challenge Attack. The configurations and incumbent numbers to beat, all taken from the file, are: 128-stage Arbiter PUF, where logistic regression with RProp reports a prediction rate of 99.9% in 2.10 seconds using 39,200 CRPs (Modeling attacks on physical unclonable functions); 128-bit 4-XOR, where the same method reports greater than 99% within 2 hours 52 minutes using 24,000 CRPs, and where CMA-ES converged on a 4x64 XOR PUF with 250,000 CRPs (Modeling Strong Physically Unclonable Functions with Metaheuristics); 16-XOR simulated perfectly reliable, where the chosen-challenge CMA-ES attack reports 93.6%–94.8% accuracy and a 21-minute compromise, against 94 minutes for the Becker attack (Breaking XOR Arbiter PUFs without Reliability Information); 64-bit 8 feed-forward-loop, where Evolution Strategies report 95%–97% prediction with 50,000 CRPs and 27 hours 20 minutes (Modeling attacks on physical unclonable functions); and $(k_\text{up}, k_\text{down})$-Interpose PUFs up to $(8,8)$ with 64-bit challenges, where the splitting attack reports accuracies above 95% but needs up to 300 million CRPs for $(8,8)$ and 750 million for $(1,9)$ (Splitting the Interpose PUF: A Novel Modeling Attack Strategy). Metrics are held-out prediction accuracy, CRPs consumed to reach 95% and 99%, and wall-clock time. Hardware validation targets the reported regime of above 96% accuracy for up to 9-XOR configurations on FPGA (Breaking XOR Arbiter PUFs With Chosen Challenge Attack). The headline targets are the two where incumbent costs are extreme: cutting the $(1,9)$-iPUF CRP requirement from 750 million toward the tens of millions, and beating 21 minutes on 16-XOR at matched accuracy with a plain supervised attack that needs no reliability or chosen-challenge access.
Baselines
The naive reference point is random guessing at 50% accuracy, the floor every attack implicitly beats; on obfuscated designs any method stuck near it is failing. The strongest incumbent overall is the CMA-ES-based chosen-challenge attack of Breaking XOR Arbiter PUFs without Reliability Information: it works on noise-free high-XOR designs where reliability-based attacks fail, reports 93.6%–94.8% on simulated perfectly reliable 16-XOR APUFs with a 21-minute compromise, and its hardware companion reports above 96% up to 9-XOR on Xilinx FPGA (Breaking XOR Arbiter PUFs With Chosen Challenge Attack). It will be run on the same synthetic generators, seeded identically, with the same throttled CRP stream and an equal tuning budget over its population and covariance parameters; beating it means reaching its accuracy band with at most its wall-clock time and strictly fewer CRPs — and, to show the regimes differ, also running it on iPUF and feed-forward configurations it was not designed for. The second baseline is the splitting attack on Interpose PUFs (Splitting the Interpose PUF: A Novel Modeling Attack Strategy), the only published attack on that family, run with its own iterative LR procedure on the same $(1,9)$ and $(8,8)$ instances; beating it means above 95% accuracy with a CRP budget at least an order of magnitude below its 300–750 million. The third baseline is logistic regression with RProp (Modeling attacks on physical unclonable functions), which should win on plain Arbiter PUFs and lose on feed-forward and interposed ones — that asymmetry is itself a result, since it maps exactly where threshold-crossing gradients add value. Finally, the ablation that isolates my key ingredient is the same architecture with the surrogate backward pass removed: either a fully smooth forward model (sigmoid in place of every step function, which collapses the method toward plain logistic regression on the composition) or a hard forward model trained by CMA-ES instead of gradients. If the fully smooth variant matches the surrogate version, the win comes from the structural decomposition, not from differentiable spike approximation, and I will report that honestly — the decisive check is comparing all three on 8-XOR and $(1,9)$-iPUF at matched CRP budgets.
Why not just…?
Why not just use logistic regression with RProp? Because the file's own evidence marks where it stops: it reports 99.9% on a 128-stage Arbiter PUF and 99% on 4-XOR in under three hours, but it could not handle Feed-Forward Arbiter PUFs at all, forcing the original authors to Evolution Strategies at 27 hours for 95–97% (Modeling attacks on physical unclonable functions), and the Interpose attack needed to bolt a bit-guessing heuristic onto it, at up to 750 million CRPs (Splitting the Interpose PUF: A Novel Modeling Attack Strategy). Logistic regression cannot pass error through an intermediate threshold, which is precisely what surrogate gradients do; the feed-forward and interpose experiments are designed to show this gap directly.
Why not just use CMA-ES? It is the strongest real competitor and I treat it as the primary baseline rather than dismissing it. Its documented costs are the opportunity: 250,000 CRPs for a 4x64 XOR PUF (Modeling Strong Physically Unclonable Functions with Metaheuristics), and a projected 1875 years to reach 60% against the trinary quadruple response design (Reliable and Modeling Attack Resistant Authentication of Arbiter PUF in FPGA Implementation With Trinary Quadruple Response). CMA-ES spends CRPs to estimate a covariance matrix it could instead compute exactly from a gradient; if the surrogate attack does not beat it on CRPs-to-accuracy at equal tuning budget on 8-XOR and above, the hypothesis is wrong and the experiments will say so.
Why not just train a standard smooth neural network (an MLP)? This is the most serious alternative, because a sigmoid MLP is itself a smooth relaxation of threshold circuits, and a standard DNN has already served as a software clone during enrollment in Reliable and Modeling Attack Resistant Authentication of Arbiter PUF in FPGA Implementation With Trinary Quadruple Response. Two things distinguish the proposal, and one is checkable in the plan itself. Structurally, my clone is not a generic approximator but the PUF's own wiring with a handful of parameters per chain, so it should need far fewer CRPs than a free-form MLP at equal accuracy — a head-to-head MLP baseline on 8-XOR and $(1,9)$-iPUF will quantify this. Methodologically, the surrogate design keeps the hard step in the forward pass, so the trained model is a faithful discrete clone rather than a probability estimator, and it avoids the compounding smooth-relaxation error through stacked XOR and interpose layers that SURGE: Surrogate Gradient Adaptation in Binary Neural Networks showed zeroes out informative gradients in binary compositions. If the MLP matches the surrogate clone's CRP efficiency, that is a genuine negative result for the spike-approximation ingredient and I will frame the contribution accordingly.
Decisive Checks
The proposal stands or falls on three comparisons, all pre-registerable. On $(1,9)$-iPUF, does end-to-end surrogate training reach 95% with at most a tenth of the 750 million CRPs the splitting attack needs? On 16-XOR, does a plain supervised surrogate attack match the chosen-challenge attack's 93.6%–94.8% within its 21-minute budget, without reliability or chosen-challenge access? And in the ablation, does the hard-forward/surrogate-backward clone beat both its fully smooth and its gradient-free twins at matched budget? A yes on the first two with a yes on the third establishes that differentiable spike approximation is a practical attack tool, not just a neuromorphic training trick; a no on the third, with yes on the first two, still establishes that differentiable structure — not spikiness — is what breaks the sample-complexity barrier, and would redirect this program toward where thresholds genuinely matter: the metastable, trinary-output designs that today's attacks project at millennia.
References
- Modeling attacks on physical unclonable functions
- Splitting the Interpose PUF: A Novel Modeling Attack Strategy
- Modeling Strong Physically Unclonable Functions with Metaheuristics
- Surrogate Gradient Learning in Spiking Neural Networks
- Spatio-Temporal Backpropagation for Training High-Performance Spiking Neural Networks
- Breaking XOR Arbiter PUFs With Chosen Challenge Attack
- Hybrid Macro/Micro Level Backpropagation for Training Deep Spiking Neural Networks.
- SURGE: Surrogate Gradient Adaptation in Binary Neural Networks
- Differentiable hierarchical and surrogate gradient search for spiking neural networks
- BindsNET: A Machine Learning-Oriented Spiking Neural Networks Library in Python
- Reliable and Modeling Attack Resistant Authentication of Arbiter PUF in FPGA Implementation With Trinary Quadruple Response