When a chip is fabricated by an untrusted offshore foundry, the design house withholds the top metal layers and finishes them at a trusted facility — a scheme called split manufacturing. The foundry still sees the placement of every logic gate, and because commercial place-and-route tools put connected gates near each other to save wire, that placement leaks the missing wiring: an attacker can reconnect driver pins to input pins by guessing that nearby pins belong together. Two recent defenses claim to close this hole completely, driving the attacker's correct-connection rate to 0%. This proposal argues that the 0% number is an artifact of measuring the attack at the wrong resolution — the wrong "frequency band," in a sense borrowed from neural signal analysis. Local placement optimization writes high-frequency clues into the layout (nearby gates connect); global placement writes low-frequency clues (which functional module sits where on the die, and how modules feed each other). Existing defenses suppress the first kind while leaving the second largely untouched, and no published attack measures either kind separately.
The core claim: if we project the placement-derived leakage onto the eigenbasis of the circuit's graph Laplacian — the graph analogue of splitting an EEG trace into delta, theta, and gamma bands — we can (1) quantify, per band, how much connectivity information a given defense actually destroys, and (2) build an attacker that treats a heavily defended layout as a low signal-to-noise measurement and recovers a functionally usable netlist from the surviving low-frequency bands plus a generative prior over what ordinary netlists look like. The security question for split manufacturing then becomes quantitative and band-resolved instead of binary. The stakes are the standard ones for hardware IP theft: a reconstructed netlist enables cloning, overproduction, and targeted Trojan insertion.
Prior Work and the Gap It Leaves
Three lines of work set up the problem. Attacks on split manufacturing have moved from minimum-cost network-flow matching of nearby pins to a supervised multi-modal network that scores candidate driver–sink pairs from layout snapshots and distance features, reporting up to a 1.21× improvement in Correct Connection Rate (CCR) over network-flow at a Metal-1 split and 1.12× at Metal-3, with inference in under 1% of network-flow runtime Attacking Split Manufacturing from a Deep Learning Perspective. Defenses answer in kind: netlist randomization before placement, with functionality restored in the trusted layers via correction cells, reports 0% CCR from proximity attackers, an output error rate of roughly 100%, and a Hamming Distance of about 40% between the attacker's reconstruction and true behavior, at 0% area, 3.5% power, and 2.7% delay overhead Raise Your Game for Split Manufacturing: Restoring the True Functionality Through BEOL; concerted wire lifting, which hoists high-fanout nets into the trusted layers, reports 0% CCR on several benchmarks and an average netlist recovery of 31% versus about 89% in prior art, at 0.85% power and 0.83% delay cost Concerted Wire Lifting: Enabling Secure and Cost-Effective Split Manufacturing. A third paper makes leakage itself measurable, modeling it as mutual information between physical distance and connectivity and cutting attack success by 5.27× (M1), 3.19× (M2), and 1.73× (M3) using graph coloring and gate-type clustering, at roughly 60% area and under 18% delay overhead Rethinking split manufacturing: An information-theoretic approach with secure layout techniques.
The unresolved tension is explicit in the corpus: the deep-learning attack was never evaluated against the randomized or wire-lifted layouts — no reported result found on Lacuna for the deep-learning attack under any published defense — and every defense's 0% CCR is measured only against proximity matching, whether heuristic or learned, that reads the layout as a flat spatial map. Nobody has decomposed the leakage signal itself. The methodological tools come from the other two directions. Frequency-band analysis of neural signals shows that broadband averages cancel opposing effects: motor movement simultaneously decreases low-frequency power and increases gamma power, so a broadband readout sees nothing Spectral Changes in Cortical Surface Potentials during Motor Movement, and distinct bands carry non-redundant information about magnitude versus timing Different LFP frequency bands convey complementary information about the BOLD signal. The low-SNR literature supplies the recovery recipe: when raw-input reconstruction fails because the loss chases noise, predict in latent space instead EEG2Rep: Enhancing Self-supervised EEG Representation Through Informative Masked Inputs, and when conditioning is extremely noisy, denoise the embedding with a generative prior before decoding rather than mapping directly to the output Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion. An explicit frequency encoder is not optional seasoning: Brant's ablations show that removing it significantly degrades performance, and its band-aware encoding improved seizure-detection F2 by nearly 30% over the previous best pre-trained model while needing 10× less downstream labeled data Brant: Foundation Model for Intracranial Neural Signal. We transfer this entire recipe, but the "signal" is the layout's spatial-affinity field and the "noise" is adversarial obfuscation rather than recording artifact.
Core Claim
Define the netlist's graph Laplacian $L = U \Lambda U^\top$ from the true gate-level connectivity, and let $S$ be the affinity matrix an attacker observes, where $S_{ij}$ decays with the physical distance between gates $i$ and $j$ in the untrusted layers. Projecting $S$ onto consecutive groups of Laplacian eigenvectors splits it into graph-frequency bands. The band leakage
$$\rho_k \;=\; \frac{\langle A,\; U_k U_k^\top S\, U_k U_k^\top \rangle_F}{\|A\|_F \,\|S^{(k)}\|_F}$$
measures how much of the true adjacency $A$ survives in band $k$, where $U_k$ are the eigenvectors of band $B_k$ and $S^{(k)}$ is the banded affinity. Low bands (small eigenvalues) capture global module-scale structure; high bands capture gate-local wiring. The claim has three parts: existing defenses drive $\rho_k$ to near zero for high $k$ but leave $\rho_k$ substantial for low $k$; a 0% CCR defense is therefore compatible with large coarse-graph recovery; and an attacker that conditions a generative graph model on the surviving bands can convert that residual leakage into a netlist good enough to clone or Trojan. The spatial signal-to-noise ratio is precisely $\rho$ aggregated over bands — and defenses change its spectrum, not just its level.
Research Program
The first workstream is a measurement study, not an attack: build the leakage spectrum. Place and route the benchmark circuits (Nangate 45nm library, matching prior work), apply each defense in its published configuration, and tabulate $\rho_k$ plus a per-band mutual-information analogue following the information-theoretic setup of Rethinking split manufacturing: An information-theoretic approach with secure layout techniques. The deliverable is a band-by-band table of what wire lifting, netlist randomization, and g-color/g-type each actually destroy, plus a formal statement of the boundary — the eigenvalue cutoff below which leakage persists. My working hypothesis, motivated by how the defenses work: wire lifting targets high-fanout nets and short-wire camouflage, i.e., local hints, while global placement still groups true functional modules because the EDA engine's partitioning step is untouched.
The second workstream is the band-aware attacker, built in the low-SNR mold. Pre-train a graph autoencoder on the corpus of open benchmark netlists to learn a latent space of plausible circuit topology, mirroring how the EEG decoding pipeline uses a pre-trained generative prior to clean noisy embeddings before decoding Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion. At attack time, encode the defended FEOL as band-weighted spectral descriptors of $S$ (a layout "power spectrum" over graph frequency, in the spirit of Brant's PSD-weighted band embeddings), condition a latent denoiser on them, and decode candidate netlists that are then filtered against observable FEOL constraints such as pin positions, fanout limits, and partial lower-metal routing. Crucially, following EEG2Rep's finding that raw-input reconstruction objectives make the model fit the noise EEG2Rep: Enhancing Self-supervised EEG Representation Through Informative Masked Inputs, training predicts in latent space rather than reconstructing adjacency entries directly — the obfuscation is adversarial noise, an even harsher version of the regime where sample-level objectives fail.
The third workstream revises the security metric and closes the loop with the IP-theft application. Report, alongside CCR, a module-level recovery score (agreement of the recovered coarse graph with true partitions), a functional score (Hamming Distance and output error rate of the recovered netlist under simulation, on the same definitions used in Raise Your Game for Split Manufacturing: Restoring the True Functionality Through BEOL), and a downstream-utility score: whether the recovered netlist suffices to plan a Trojan insertion that survives a netlist-level detector in the style of TrojanSAINT, which achieves 78% true-positive and 85% true-negative rates on unseen TrustHub designs TrojanSAINT: Gate-Level Netlist Sampling-Based Inductive Learning for Hardware Trojan Detection. Finally, the leakage spectrum feeds back into defense design: a spectral version of the g-color placement constraint that also decorrelates the low bands, evaluated on the same overhead budgets, so the project produces a matched attack–defense pair rather than only a stronger attack.
Proposed Experiments
Benchmarks are fixed by the prior literature so results are directly comparable: ISCAS-85, MCNC, and ITC-99 circuits, plus IBM Superblue designs for industrial scale, all placed with the Nangate 45nm library — the exact suites used by the attack and defense papers above. Defenses are applied in their published configurations. The primary metrics are CCR (connection-level), Percentage of Netlist Recovery (netlist-level), module-level recovery, and functional Hamming Distance and output error rate; overhead metrics (area, power, delay) apply only to the defense side. The decisive checks are: first, the leakage-spectrum prediction — wire-lifted and randomized layouts show near-zero $\rho_k$ at high bands and materially nonzero $\rho_k$ at low bands, which if false kills the premise and I would report it as such; second, the recovery prediction — on wire-lifted designs the attacker must exceed the 31% average netlist recovery reported as the achieved security level in Concerted Wire Lifting: Enabling Secure and Cost-Effective Split Manufacturing under a matched metric definition; third, on randomized designs, where the defense reports 0% CCR and roughly 40% attacker Hamming Distance, the attacker succeeds if it cuts the functional Hamming Distance well below 40%, i.e., produces a reconstruction that misbehaves substantially less than the defense guarantees; and fourth, the ablation grid — the same latent attacker with broadband (non-band-decomposed) conditioning, with raw-adjacency reconstruction loss instead of latent prediction, and without the pre-trained prior, each isolated one at a time. The broadband-versus-banded ablation also empirically settles, in a new domain, the open tension between Brant's success with masked raw reconstruction Brant: Foundation Model for Intracranial Neural Signal and EEG2Rep's evidence that latent prediction beats raw reconstruction under noise EEG2Rep: Enhancing Self-supervised EEG Representation Through Informative Masked Inputs.
Baselines
The do-nothing reference is random candidate matching: connect each driver pin to a uniformly random eligible sink. It fixes the floor — any real attacker must beat it by a wide margin, and it makes the size of the leakage effect legible on every defended benchmark. The classical baseline is the network-flow proximity attack, which matches drivers to sinks by minimizing a Manhattan-distance cost; it is the historical standard and still the reference point in every defense paper, but it is the weakest serious competitor because Attacking Split Manufacturing from a Deep Learning Perspective already reports a 1.21× CCR advantage over it at M1 and 1.12× at M3, with over 100× faster inference, and because distance-based matching by construction cannot see structure once distance is decorrelated from connectivity. The strongest existing method is that deep-learning attack itself Attacking Split Manufacturing from a Deep Learning Perspective, which scores candidate driver–sink pairs from fused image and vector features; it is the competitor to beat because it is the published state of the art on exactly this attack surface. To make the comparison fair and honest, I will re-run it on identical layouts and candidate lists from the same proximity windows, retrain it on the defended layouts (it was never tested there — no reported result found on Lacuna for it under netlist randomization or wire lifting), give it the same GPU budget and an equal random-search tuning budget as my attacker, and additionally give it a fine-tuning budget on defended training circuits. Beating it means higher netlist-level recovery on defended layouts while remaining within a few points of its CCR on undefended layouts at M1 and M3 splits. Finally, the key ablation is my own attacker with the spectral band decomposition removed — identical latent model, identical prior, identical training data, but conditioned on a broadband affinity encoding — which isolates the one ingredient this proposal actually contributes. If the broadband version matches the banded one on defended layouts, the band story is wrong and the honest conclusion is that a generic low-SNR graph generator, not frequency analysis, does the work; I would report that outcome and it would still be a finding about the insufficiency of CCR-based security claims.
Why not just…?
Why not just fine-tune the existing deep-learning attack on defended layouts? This is the most serious alternative and it is folded into the experimental design rather than argued away. The honest answer is that it might partially work: the attack is a supervised scorer, and given labeled defended layouts it can learn whatever residual signal exists. But there is structural evidence it falls short of the generative approach: it scores candidate pairs independently and has no prior over what a coherent whole netlist looks like, so when pairwise spatial evidence is deliberately corrupted, it cannot fall back on global consistency the way a latent generative model conditioned on surviving low-frequency bands can — the same distinction that separates direct end-to-end decoding from prior-guided latent decoding in the low-SNR literature Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion. The head-to-head experiment above, with equal tuning and fine-tuning budgets, settles this empirically; if the fine-tuned deep-learning attack matches my recovery rates, the simpler method wins and the proposal's contribution reduces to the leakage-spectrum measurement study.
Why not just use the classical network-flow attack, which is mathematically clean? It is already superseded on both speed and accuracy — a 1.21× CCR gap and inference at under 1% of its runtime are reported by Attacking Split Manufacturing from a Deep Learning Perspective — and, more fundamentally, its entire cost function is a monotone function of physical distance, so any defense that decorrelates distance from connectivity (which is precisely what netlist randomization does) makes its optimum point in a random direction. It will be run as a baseline anyway, because it is the reference point every defense paper reports against.
Why not just defend harder instead of attacking better? From the defender's side, one could ignore the residual-leakage question and simply randomize placement without constraint. The literature already prices this: unconstrained randomization incurs over 600% layout overhead per Rethinking split manufacturing: An information-theoretic approach with secure layout techniques, which is commercially dead, and the whole trend of defense work has been buying security under tight power, performance, and area budgets. The proposal's contribution on this front is that the leakage spectrum tells the defender where to spend that budget: if low-band leakage is what survives, the g-color-style constraint should be extended to decorrelate module-level placement, and the final experiments test exactly that under the same overhead accounting, so the defensive alternative is answered with a construction rather than a dismissal.
Why not just use a bigger end-to-end graph network without any band decomposition? The corpus is unusually direct on this: broadband analysis cancels opposing effects Spectral Changes in Cortical Surface Potentials during Motor Movement, distinct bands carry non-redundant information Different LFP frequency bands convey complementary information about the BOLD signal, Brant's ablations show removing explicit frequency encoding significantly hurts Brant: Foundation Model for Intracranial Neural Signal, and latent prediction beats raw end-to-end training by over 4% on DREAMER in the low-SNR regime EEG2Rep: Enhancing Self-supervised EEG Representation Through Informative Masked Inputs. These are analogies, not proofs, which is why the broadband ablation is identical in every respect except the band decomposition — if capacity alone closes the gap, the banded variant adds nothing and the simpler design should be preferred.
Decisive Checks
The program succeeds if the leakage spectrum shows a clean band split under published defenses, and if the band-aware attacker exceeds 31% netlist recovery on wire-lifted benchmarks and drives functional Hamming Distance well below the ~40% that netlist randomization currently guarantees against proximity attackers. It fails informatively if defenses turn out to suppress low bands too — in which case split manufacturing's security claims hold for a reason nobody has yet articulated, and the band-resolved mutual-information measurement is itself the contribution. Either outcome replaces a binary, attack-specific security claim with a quantitative leakage model the next defense can be designed against.
References
- Attacking Split Manufacturing from a Deep Learning Perspective
- Raise your game for split manufacturing: restoring the true functionality through BEOL
- Concerted Wire Lifting: Enabling Secure and Cost-Effective Split Manufacturing
- Rethinking Split Manufacturing: An Information-Theoretic Approach with Secure Layout Techniques
- Spectral Changes in Cortical Surface Potentials during Motor Movement
- Different LFP frequency bands convey complementary information about the BOLD signal
- EEG2Rep: Enhancing Self-supervised EEG Representation Through Informative Masked Inputs
- Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion
- Brant: Foundation Model for Intracranial Neural Signal
- TrojanSAINT: Gate-Level Netlist Sampling-Based Inductive Learning for Hardware Trojan Detection