Source-linked AI summary
s-MDM: Generative Virtualization of Multi-Device Hardware Variations for Portable DL-SCA
Niloufar Sayadi, Marten van Dijk, Chenglu Jin
TL;DR
DL-SCA classifiers can degrade on unseen hardware as physical leakage distributions shift, while existing portability approaches rely on physical fleets or target traces. s-MDM synthesizes virtual device styles offline from source-device traces. It performs better on the layout/acquisition-shifted Piñata target, whereas physical MDM remains superior on the electrical-clone target D4.
Problem
Unseen-hardware shifts in process, PCB routing, and measurement noise alter leakage distributions, while physical-fleet and target-assisted portability approaches have operational limits.
Source: §1 · p. 1Method
s-MDM learns a structured generative model from profiling devices and synthesizes virtual device-style traces offline for downstream attack training, with the final target held out.
Source: §2 · pp. 1–2Results
Takeaways & Limitations
The reported operational boundary favors physical MDM when the target is an electrical clone, and s-MDM under layout and acquisition shift.
Source: §4 · p. 3
Abstract
from arXiv · showhide
Deep Learning-based Side-Channel Analysis (DL-SCA) frequently suffers from catastrophic performance degradation across unseen hardware due to printed circuit board routing differences, silicon process variations, and measurement noise shifts. This poster presents the Synthetic Multiple Device Model (s-MDM), a zero-target-trace generative framework designed to improve cross-device portability. s-MDM combines a structured cVAE generator, a Walsh-Hadamard leakage anchor, continuous style modulation, and decoupled leakage-style--domain critics to synthesize virtual source-device profiles offline. Benchmarked on 32-bit side-channel traces (AES_PTv2), s-MDM maps a precise operational boundary: while physical MDM remains superior on identical electrical clones (D4), s-MDM achieves consistently low key rank on the layout/acquisition-shifted Pinata target, where physical baselines are unstable or misaligned.
1 Introduction
DL-SCA classifiers can lose performance on unseen hardware because process variations, PCB routing differences, and noise shifts alter physical leakage distributions. Existing portability approaches either require costly physical fleets or target-device traces, while s-MDM synthesizes virtual device styles from source traces alone.
1 Introduction
- Unseen-hardware portability is difficult because process variations, PCB routing differences, and noise shifts alter physical leakage distributions.These shifts can severely degrade classifier performance.Source: §1 · p. 1
- Physical MDM needs a fleet of boards, while target-assisted adaptation requires unlabeled target traces during profiling.Fleet procurement and profiling carry financial and logistical overheads; target-trace requirements can be infeasible in non-permissive scenarios.Source: §1 · p. 1
- s-MDM uses source-device traces to synthesize a continuous device-style manifold and examines when physical pooling or synthetic virtualization generalizes better.Its hybrid cVAE-GAN combines a leakage anchor, continuous style modulation, decoupled critics, and Gradient Reversal Layers.Source: §1 · p. 1
2 Methodology
s-MDM trains a structured generative model on available profiling devices, then synthesizes varied virtual-device traces offline for downstream attack-classifier training. A held-out source-family board supports validation while the final target remains unseen.
2 Methodology
- s-MDM uses only available profiling-device traces to learn a generator, then synthesizes virtual device styles offline instead of adapting with target traces.The synthesized dataset trains a downstream attack classifier; the generator and critics are discarded before key recovery.Source: §2 · pp. 1–2
- The trace decomposition separates deterministic AES-dependent leakage, stochastic content variation, and device-specific style represented by a continuous 16-dimensional code.The deterministic template uses a 256-dimensional Walsh–Hadamard basis representation of the 8-bit intermediate value.Source: §2 · p. 2
- A structured cVAE generator estimates content and residual variance while keeping the deterministic leakage template anchored during optimization.This design couples cVAE density modeling with adversarial feedback to address VAE waveform blurring.Source: §2 · p. 2
- Continuous style modulation re-renders traces by modulating residual content noise with a style code through FiLM.The style term is implemented using trace-length scaling and shift vectors.Source: §2 · p. 2
- Decoupled critics evaluate physical trace morphology and enforce cryptographic alignment, while a GRL domain head discourages board-identifying features.The leakage critic preserves trace–label alignment; the domain head predicts profiling-board identity from the feature map.Source: §2 · p. 2
- The generator samples style codes from an expanded latent envelope to interpolate and mildly extrapolate beyond observed source-board appearances.The style code is sampled from a normal distribution with α = 1.5, without using final-target traces.Source: §2 · p. 2
- LOBO validation holds out a source-family board as a pseudotarget, keeping the final target unseen during training and validation.The held-out board supports checkpoint selection and leakage-fidelity monitoring.Source: §2 · p. 2
3 Experimental Evaluations
Evaluation maps an operational boundary: s-MDM performs well on the layout/acquisition-shifted Piñata target, while physical MDM remains preferable on the electrical clone D4. Ablations identify continuous style diversity as the only component whose removal produces a resolved degradation.
Experimental Setup
- D1/D2 profile, D3 validates, and D4/Piñata test clone versus layout/acquisition shifts; ρ is 0.072 on D1 and 0.737 on Piñata.Traces are cropped around each device’s leakage peak to a common 120-sample width; Table 1 characterizes the fleet.Source: §3 · p. 2, Table 1 · p. 2
In-Manifold Evaluation
- On electrical clone D4, Physical MDM reaches full recovery in 109 traces versus 586 for the two-device baseline; s-MDM does not converge within the evaluated budget.The results are consistent with real multi-device traces being more informative than synthetic re-rendering in the clone setting.Source: §3 · pp. 2–3, Table 2 · p. 3
Out-of-Manifold Evaluation
- Across Piñata’s three realizations, s-MDM achieves GE ≤3.3 while physical baselines are unstable, with Physical MDM ranking worse than random.The target has SNR = 2.796 and ρPiñata = 0.737.Source: Table 3 · p. 3, §3 · p. 3
Ablation Study
- Only removing continuous style diversity degrades beyond the full model’s seed spread (GE = 7.39; SR falls from 0.71 to 0.33); remaining ablations are unresolved.Each ablation is a single run, and the full configuration spans GE = 0.42–3.28 across three seeds.Source: §3 · p. 3, Table 4 · p. 3
4 Conclusion and Future Work
s-MDM improves DL-SCA portability under severe layout and acquisition shifts, but its advantage has a boundary: physical MDM remains superior on electrical clones. Continuous style diversity is the only ablated component whose removal causes a resolved degradation; critic roles require multi-seed replication.
Conclusion
- s-MDM improves DL-SCA portability under severe layout and acquisition shifts, while physical MDM remains superior on electrical clone D4.On Piñata, physical baselines are unstable or anti-correlated.Source: §4 · p. 3
- Removing continuous style diversity is the only ablation that causes a resolved performance degradation; critic roles require multi-seed replication.Future work will refine critic optimization, calibrate style priors, and evaluate larger multi-board fleets.Source: §4 · p. 3
- The reported scope is AES_PTv2, with a four-board STM32 fleet and a distinct Piñata target.The conclusion calls for evaluation on larger multi-board fleets.Source: §4 · p. 3