Brain imaging solved a problem that hardware security has not. Every brain is different, yet neuroscientists routinely find a few cubic millimetres of abnormal tissue. They do it with voxel-based morphometry (VBM). Each scan is warped onto a shared template, sorted into tissue types, and blurred slightly. A separate statistical test is then run at every 3D pixel ("voxel"), comparing that location against how much it normally varies across a population. Chip assurance faces the same problem with different materials. An untrusted foundry can add a few cells in empty layout space, swap in a slower cell, or reroute a wire. Every manufactured die also differs from its design through ordinary process variation. Current physical inspection either compares 2D microscope images to the design with whole-image overlap scores or relies on manual inspection. Neither has a principled way to say "this spot deviates more than manufacturing noise allows."
We propose a direct translation of the VBM statistical machinery to 3D scans of manufactured chips. Each die volume is registered to a voxel rendering of its own design layout and segmented into materials (silicon, polysilicon, metal, vias). Each voxel is then tested against a variance map estimated from a population of dies. The test's spatial scale is chosen from a taxonomy of physical attack footprints, so the taxonomy predicts which tamper types are detectable, at what size, and with which settings. The bet is narrow and testable: per-voxel population variance plus footprint-matched smoothing should beat plain "scan minus design, then threshold" at a fixed whole-die false-alarm rate. The gain should appear specifically where process variation is spatially uneven.
Core Claim
The claim has two parts. The first is a detection claim. A voxel statistic that normalizes each location by its own population-estimated process variance should detect tamper footprints at lower size thresholds than a golden-difference baseline using one global noise level. Family-wise error control should also make the whole-die false-alarm rate a stated design parameter rather than a tuned threshold. Variance concentrates at metal edges and in dense routing, so a global threshold is either blind in quiet regions or noisy at edges.
The second is a categorization claim. Physical attack vectors should be classified by their volumetric footprint: which material compartment they change, their spatial extent, and whether they add, remove, or displace material. These three properties predict detectability better than the logical categories (trigger type, payload, RTL vs. gate level) used in the Trojan literature. Some classes should be provably out of reach of this imaging-based test. Dopant-only changes are the obvious case, and the taxonomy should mark them as such rather than hide them.
Let $S_d(v)$ be the smoothed material map of die $d$ at voxel $v$ after registration to the design template. Let $\mu(v)$ be the expected value there, rendered from the design layout through a process model. Let $\sigma(v)$ be the standard deviation of $S(v)$ across a reference population of registered dies. The per-voxel statistic is
$$ z_d(v) = \frac{S_d(v) - \mu(v)}{\sigma(v)}. $$
The die is flagged if $\max_v |z_d(v)|$ exceeds a family-wise threshold. For smoothed fields, that threshold grows only roughly as $\sqrt{2\log N}$ in the number of independent resolution elements $N$, a standard Gaussian-extreme-value result that random field corrections refine. Billion-voxel dies therefore remain testable. Using the design for $\mu$ and the population only for $\sigma$ is deliberate. A mask-level Trojan inserted at the foundry appears in every die, so a die-versus-population test would absorb it. A die-versus-design test with population variance would not.
Positioning Against Prior Work
The method comes from neuroimaging. Mapping Brain Tissue Volume with Voxel-Based Morphometry sets out the normalize–segment–smooth–test pipeline. Mapping Brain Tissue Volume with Voxel-Based Morphometry shows that voxel-wise testing finds subtle focal abnormalities that expert visual inspection misses. That is the property we want for tamper footprints. Confirming the Diversity of the Brain after Normalization shows that nonlinear diffeomorphic warping keeps fine individual detail, with 100% identification and 0.072% equal error rate across 150 OASIS subjects scanned over a year apart. For us this cuts both ways. It suggests that warping does not erase small differences. It also raises the worry that an overly flexible warp could bend a die's anomaly into "normal" shape. We treat warp flexibility as an ablated variable rather than a default.
The hardware motivation comes from three findings. REFICS found two problems with 2D microscope (SEM) pipelines. Models fail across layers and technology nodes, and IoU/SSIM do not track electrical integrity, since a high-overlap segmentation can hide a broken trace. Per-voxel significance maps answer that metric problem directly, because they localize rather than average. Histogram-based Auto Segmentation supplies a parameter-free material segmentation we reuse for the segmentation stage. It stops at 2D segmentation and does no statistical testing. FICS PCB X-ray supplies the only real 3D X-ray CT volumes in our file, and analyses them only descriptively.
The constraining work is logic-level detection. HaTCh detects all TrustHub, DeTrust, and k-XOR-LFSR Trojans at 4.18% area overhead. It applies only to trigger-activated digital logic and needs on-chip circuitry. TrojanSAINT and TROJAN-GUARD learn on netlist and RTL graphs. Both assume the netlist is the artifact under suspicion. Neither sees foundry-introduced physical changes that leave the delivered netlist intact. Securing Integrated Circuits Against Untrusted Foundries describes exactly such changes, for example substituting functionally equivalent slower cells.
The categorization lens comes from Taxonomy-Driven Graph-Theoretic Framework for Manufacturing Cybersecurity Risk Modeling, which ties attack vectors to physical locations and consequences at factory scale. We push that location-attribute idea down to the die. We also replace heuristic success probabilities with measured detection probabilities per footprint class. That addresses one of that framework's acknowledged weaknesses.
Research Program
We first build the footprint taxonomy. Physical attack vectors from the cited hardware literature are mapped to four kinds of change:
- added cells in white space;
- cell substitution of equal footprint but different internal geometry;
- metal rerouting or via removal, which are BEOL edits in the sense of Raise Your Game for Split Manufacturing;
- dopant-profile changes.
Each class is annotated with the compartment it touches, its extent in voxels at a given scan resolution, and whether it adds or removes material. This annotation fixes the smoothing kernel for each class, following the matched-filter principle that blurring at the signal's own scale maximizes signal-to-noise. It also fixes which statistic is used. Added material should appear in the intensity residual. Displaced material should appear in the Jacobian determinant of the warp, the VBM "modulation" signal measuring local stretching.
Next we build the pipeline. A design layout is rendered to a voxel template per material. Each die volume is registered to it in three increasing levels of warp freedom: rigid, affine, and constrained nonlinear. The die is segmented into materials, smoothed per footprint class, and tested as above. Family-wise error is controlled across the whole die, with false-discovery-rate control as a secondary option.
The data are mostly simulated, which is the main practical limitation. There is no public 3D silicon CT dataset in our file. We therefore render ISCAS-85 and IBM Superblue layouts (Nangate 45nm, as in the split-manufacturing work) and ISPD 2015/2019 placements into voxel volumes. Populations of 50–500 dies are generated with a process-variation model: line-edge roughness, layer thickness drift, and density-dependent dishing. Imaging noise follows REFICS's physics-based SEM noise model, extended across stacked layers at the 32/28nm and 90nm settings. Tamper footprints are injected per taxonomy class at controlled sizes.
On the real side, the FICS PCB X-ray volumes provide a check on noise and registration at PCB scale. PCB 1 has 13.59 μm resolution, 1,430 slices, and 164 annotated vias. There we digitally remove vias and add or reroute traces in real reconstructed volumes, with the scan's own repeat structure serving as the population.
Baselines
The naive reference is golden differencing. The die volume is registered to the rendered design with the same registration as our method, the absolute material-map difference is computed, and it is thresholded with one global threshold. That threshold is calibrated on clean dies to the same die-level false-alarm rate (0.05) we use. This is what a careful engineer would do first, and it isolates our two ingredients by lacking both.
Two ablations isolate the key ingredient. The first is our full statistic with a global $\sigma$ in place of $\sigma(v)$. If this matches the full method, the population-variance idea adds nothing. The second is our statistic with a single fixed smoothing kernel in place of footprint-matched kernels. If this matches, the taxonomy adds nothing for detection.
The strongest learned competitor is a 3D U-Net segmentation network trained to mark tampered voxels. It is trained on the same simulated populations with injected footprints from a held-out subset of taxonomy classes, under an equal compute and tuning budget. Its die-level threshold is calibrated to the same false-alarm rate. It is the strongest competitor because supervised detectors usually win when training and test distributions match. The honest question is whether it holds up on unseen footprint classes and on a different node, where REFICS reports qualitatively severe cross-node drops.
For reference only, we report the netlist-level numbers on a different input. TrojanSAINT reports 78% true positive rate and 85% true negative rate on unseen TrustHub circuits, and HaTCh reports 100% detection on TrustHub. We found no reported result on Lacuna for detecting physical tampering in 3D chip or PCB volumes, so there is no published number on our exact task to beat.
Beating the baselines means two things. First, a lower minimum detectable footprint size than golden differencing at equal die-level false alarm, in at least the edge-dense metal classes. Second, equal or better detection than the U-Net on footprint classes and nodes it was not trained on.
Why not just…?
Why not just subtract the scan from the design and threshold? This is the real threat. If process variation were spatially uniform, our statistic would reduce to exactly this. The proposal stands or falls on the global-$\sigma$ ablation. If per-voxel variance gives no gain on dense metal, we will report that VBM's contribution here is only family-wise error calibration and smoothing. That is a much weaker claim.
Why not just train a 3D detector on simulated Trojans? A supervised network will likely win on the footprint classes it saw. REFICS documents that vision models trained on one layer or node fail on another, and a statistic that needs no tamper examples cannot overfit to tamper shapes. The held-out-class and cross-node splits test exactly this. If the U-Net generalizes, the case for our method narrows to calibrated false-alarm control and interpretability.
Why not just use HaTCh or netlist graph networks? They answer a different question: whether the logic is malicious. They cannot see foundry edits that leave the delivered netlist intact, such as equal-footprint cell substitution or BEOL rerouting. HaTCh also costs 4.18% area and covers only triggered digital logic. These methods complement ours. Our taxonomy table will state which classes each method covers.
Why not rely on supply-chain controls instead of algorithms? You Don't Need Robust Machine Learning to Manage Adversarial Attack Risks shows that process controls beat model hardening for black- and grey-box software attackers. A foundry has full physical access to the masks and wafers, so perimeter controls do not reach the artifact. Inspecting the artifact itself is the remaining check.
Proposed Experiments
The first decisive check is a detectability curve per taxonomy class. For each class we plot detection probability against footprint size at die-level family-wise false alarm of 0.05, for the full method, both ablations, golden differencing, and the U-Net. The curves are reported separately for low- and high-variance regions. The predicted signature is a gap between full and global-$\sigma$ curves that appears in high-variance regions and vanishes in low-variance ones. If the gap does not appear in that pattern, the core claim fails.
The second check addresses whether warping absorbs the footprint. We repeat the curves under rigid, affine, and nonlinear registration and test whether the Jacobian statistic recovers what nonlinear warping removes from the intensity residual. This tests the tension raised by the brain-identification result.
The third check is calibration. On clean simulated dies and untampered FICS PCB volumes, the observed die-level false-alarm rate should match the nominal 0.05 within sampling error. A miscalibrated test would be no better than a tuned threshold.
The fourth check is connectivity. Flagged voxel clusters are scored by whether they intersect a net whose electrical connectivity changed, rather than by IoU, following REFICS's metric critique.
The fifth check is population size. We vary the reference population from 10 to 500 dies to find how many are needed before $\sigma(v)$ stabilizes, which bounds practical cost.
We will also report the classes the method cannot detect, dopant-only changes in particular, as negative results within the taxonomy.
References
- Regional Abnormality of Grey Matter in Schizophrenia: Effect from the Illness or Treatment?
- Objective Detection of Epileptic Foci by18F-FDG PET in Children Undergoing Epilepsy Surgery
- Confirming the Diversity of the Brain after Normalization: An Approach Based on Identity Authentication
- REFICS: A Step Towards Linking Vision with Hardware Assurance
- Histogram-based Auto Segmentation: A Novel Approach to Segmenting Integrated Circuit Structures from SEM Images
- FICS PCB X-ray: A dataset for automated printed circuit board inter-layers inspection
- HaTCh: Hardware Trojan Catcher
- TrojanSAINT: Gate-Level Netlist Sampling-Based Inductive Learning for Hardware Trojan Detection
- TROJAN-GUARD: Hardware Trojans Detection Using GNN in RTL Designs
- RTL-Breaker: Assessing the Security of LLMs Against Backdoor Attacks on HDL Code Generation
- Taxonomy-Driven Graph-Theoretic Framework for Manufacturing Cybersecurity Risk Modeling and Assessment
- Raise your game for split manufacturing: restoring the true functionality through BEOL
- You Don’t Need Robust Machine Learning to Manage Adversarial Attack Risks