---
title: 'DEMIST: Frameworks in Hardware Security & Imaging'
url: https://www.emergentmind.com/topics/demist
type: topic
---

# DEMIST: Frameworks in Hardware Security & Imaging

DEMIST refers to several distinct frameworks in the domains of hardware security and medical imaging, each with specialized methodology and application scope. In hardware security, DeMiST (Detection and Mitigation of Stealthy analog hardware Trojans) targets the detection and neutralization of highly stealthy analog hardware trojans in integrated circuits. In medical imaging, DEMIST designates both a detection-task-specific deep learning approach for denoising myocardial perfusion SPECT images, and a decoupled multi-stream latent diffusion framework to synthesize quantitative myelin maps for MRI-based assessment of multiple sclerosis. The following sections provide detailed technical expositions of these three instantiations, elucidating their theoretical foundations, system architectures, algorithmic innovations, evaluation metrics, and documented empirical performance.

## 1. DeMiST for Detection and Mitigation of Stealthy Analog Hardware Trojans

### Definition and Problem Context

A capacitance-based analog hardware trojan (AHT) exploits a minimal on-chip capacitor to accumulate charge via rare or adversarial toggling events. These trojans present with negligible area and power overhead, and can elude detection through classic side-channel analyses, functional tests, and EM localization. The DeMiST methodology introduces the first comprehensive framework for both detection and real-time neutralization of this class of threats [2310.03994].

### Charge Sharing and Accumulation Model

The crucial step in analog trojan triggering is per-toggle voltage accumulation, modeled as:
\[
V_{1} = \frac{C_\text{unit} V_{DD} + C_\text{main} V_0}{C_\text{unit} + C_\text{main}}
\]
\[
\Delta V = \frac{C_\text{unit} (V_{DD} - V_0)}{C_\text{unit} + C_\text{main}}
\]
Accumulated over $N$ toggles:
\[
V_N \approx V_0 + N \Delta V
\]
For standard on-chip parasitic values, $\Delta V$ is $\mathcal{O}(\mu V - mV)$ and $Q_N = C_\text{main} N\Delta V$ may go unnoticed unless $N$ is very large.

### Existing Detection Techniques and the "Fortified" Trojan

Conventional detection approaches—power/EM profiling, rare-event logic vectors, laser-based imaging, and switching activity localization—fail when faced with microvolt-scale, million-cycle trojan triggers. Advanced variants of the attack, such as the "fortified AHT," introduce a supplementary $V_{DD}$-gated PMOS with capacitance $C_\text{new}$ to retain charge even during power gating attempts:
\[
\Delta V = \frac{C_\text{unit} (V_{DD} - V_0)}{C_\text{unit} + C_\text{main} + C_\text{new}}
\]
This architectural modification hardens the AHT against detection via rapid $V_{DD}$ cycling.

### Two-Phase DeMiST Methodology

#### Phase 1: Run-Time Mitigation

DeMiST employs dynamic $V_{DD}$ gating: a PMOS in the power distribution network, controlled by a state machine, periodically disables supply voltage for $T_\text{off}$ out of every $T_\text{on} + T_\text{off}$ cycle. This discharges abnormal charge accumulation without interrupting correct pipeline operation, provided $T_\text{off}$ meets the leak-down threshold:
\[
T_{\mathrm{off}} \geq \tau_\text{leak} \ln\left(\frac{V_\text{threshold}}{V_{N-1}}\right)
\]

#### Phase 2: Duty-Cycle Driven Detection

In pre-market validation, the supply duty cycle $D$ is intentionally reduced while monitoring side-channel signatures. The trojan, if present, is forced to activate under these controlled conditions, resulting in observable power transients. The detection protocol uses coarse-to-fine stepping of $D$ and monitors for side-channel anomalies:
```python
for D in [100%,80%,60%,40%,20%]:
    configure_PDNGating(duty_cycle = D)
    apply_test_vectors()
    power_trace = measure_power_profile()
    if detect_spike(power_trace):
        report Trojan_detected at duty cycle D
        break
else:
    report No_Trojan_detected
```

### Implementation and Empirical Results

In 45 nm PDK and ISCAS-85 benchmarks, DeMiST yields area overhead ranging from 63.4% (s298, 75 gates) down to 1.65% (c880, 383 gates), with <2% throughput reduction in mitigation mode. All injected AHTs were detected (100% true positive rate; 0% false positive rate on 50 clean chips). Power overhead at test time is <1% [2310.03994].

### Limitations and Potential Directions

DeMiST requires integration-stage intervention (PDN gating and latches), careful calibration for process variations, and potentially higher $T_\text{off}$ for trojans with extreme retention. Future lines of research include adaptive duty cycling, machine-learning-based power analysis, extension to RF/SerDes analog blocks, and integration with EM localization.

## 2. DEMIST for Task-Specific Denoising in Myocardial Perfusion SPECT

### Clinical Motivation and Mathematical Foundation

Reducing radiation dose or acquisition time in MPI-SPECT exacerbates noise and degrades defect-detection. Conventional DL denoisers optimized on voxel-wise metrics (MSE, RMSE, SSIM) can inadvertently deteriorate task-specific detection performance, as measured by model observers (e.g. channelized Hotelling observer, CHO). DEMIST (Detection-task-specific deep-learning-based approach) was developed to simultaneously suppress noise and preserve the frequency-domain features most predictive for defect detection [2306.04249].

### Architecture and Loss Formulation

DEMIST employs a 3D U-Net, mapping low-dose reconstructions $\hat f_{LD}$ to predicted normal-dose images $\hat f_{ND}^{pred}$. The loss functional is:
\[
L(\theta) = L_\text{fid}(\theta) + \lambda L_\text{task}(\theta)
\]
where $L_\text{fid}$ is voxelwise MSE and $L_\text{task}$ is a channelized feature loss based on deviation in anthropomorphic channel outputs (frequency profiles centered on defect locations).

CHO-based evaluation is framed as:
\[
\mathrm{SNR}^2 = \Delta \bar{v}^T K_v^{-1} \Delta \bar{v}
\]
where $\Delta \bar{v}$ represents the mean difference in channel outputs and $K_v$ the covariance under different defect states.

### Training Data and Evaluation Protocol

A cohort of 338 normal MPI-SPECT studies was processed, using binomial thinning to simulate 6.25%, 12.5%, and 25% dose levels. Synthetic defects spanning diverse extents/severities were injected for controlled evaluation. Cross-validation was performed for all network hyperparameters, and model-observer AUC was the primary endpoint.

### Quantitative Results

DEMIST systematically outperformed task-agnostic denoisers and low-dose reconstructions in observer AUC, especially at the lowest dose levels.

| Dose Level | Low-dose | TADL (task-agnostic) | DEMIST | Normal-dose |
|------------|----------|----------------------|--------|-------------|
| 6.25%      | 0.60     | 0.61                 | 0.72   | 0.84        |
| 12.5%      | 0.68     | 0.70                 | 0.82   | 0.90        |
| 25%        | 0.78     | 0.82                 | 0.86   | 0.94        |

RMSE and SSIM improved moderately over low-dose and on par or slightly below TADL, but model-observer performance gains were substantial (statistically significant at $p<0.05$, Bonferroni-corrected).

### Theoretical Analysis

Channel-space SNR decomposition shows DEMIST reduces noise variance (eigenvalues $\gamma_m$) and, unlike TADL, preserves task-relevant signal coefficients ($\alpha_m$) along critical eigenmodes—leading to net gains in AUC.

## 3. DEMIST: Decoupled Multi-Stream Latent Diffusion for Quantitative Myelin Map Synthesis

### Rationale and Framework

Quantitative magnetization transfer (qMT) imaging provides pool size ratio (PSR) maps, but requires lengthy acquisition. DEMIST synthesizes PSR maps from standard T1w+FLAIR images using a two-stage 3D latent diffusion pipeline with decoupled conditioning [2511.12396].

### Two-Stage Architecture

**Stage 1:** Separate 3D AutoencoderKL models for cond (input) and PSR (target) images, yielding aligned latents $z_c$ and $z_y$:
\[
\mathcal{L}_{AE} = \|u - \hat u\|_1 + \lambda_{KL} \mathrm{KL}\bigl[q(z|u)\|\mathcal N(0,I)\bigr] + \lambda_{edge} \|\nabla \hat u - \nabla u\|_1 + \lambda_{adv} \mathcal{L}_{GAN}(\hat u, u)
\]
Alignment loss is imposed between $z_c$ and $z_y$.

**Stage 2:** Conditional latent diffusion on $z_y$, with only the conditioning modules $\mathcal{C}(z_c)$ being trainable:
\[
\mathcal{L}_{diff} = \mathbb{E}_{t,\varepsilon} \|\varepsilon - \epsilon_\theta(z_t, t; \mathcal{C}(z_c))\|_2^2 + \lambda_{edge} \|\nabla D_{PSR}(\hat z) - \nabla y\|_1 + \lambda_{align} \|z_c - z_y\|_2^2
\]

### Multi-Stream Conditioning

- **Semantic stream:** cross-attention tokens provide high-level anatomical and intensity information.
- **Spatial stream:** ControlNet-style residuals correct spatial feature misalignment.
- **Adaptive stream:** LoRA-modulated low-rank updates offer lightweight adaptation of frozen pretrained attention layers.

Parameter efficiency is maximized—only about 16 M parameters (15%) are trainable, with the ~100 M backbone frozen.

### Results and Empirical Assessment

DEMIST outperformed prior VAE, GAN, and diffusion baselines in PSNR, SSIM, and MSE, as well as ROC-AUC for lesion-vs-NAWM (normal-appearing white matter) discrimination.

| Method           | PSNR     | SSIM     | MSE         |
|------------------|----------|----------|-------------|
| DEMIST           | 25.73    | 0.8867   | 1.752e-3    |
| cWDM (best prior)| 24.49    | 0.8746   | 2.474e-3    |

Qualitatively, DEMIST produced sharper lesion boundaries and more accurate deep nuclei, with ControlNet and LoRA providing complementary spatial/semantic support.

### Future Directions

Prospective work includes multi-site generalization, efficient deterministic inference, and interactive or user-guided synthesis.

## 4. Comparison and Context Across Application Domains

Although sharing an acronym, DEMIST methodologies address fundamentally different scientific and engineering challenges: adversarial circuit integrity and deep clinical imaging. Common themes include decoupling of roles in complex architectures (e.g., semantic vs. spatial vs. adaptive streams in diffusion models; run-time vs. test-time phases in hardware security) and an emphasis on preserving or revealing critical task-relevant features that are otherwise suppressed or obfuscated by noise, measurement protocol, or adversarial design.

## 5. References to Primary Research

- "DeMiST: Detection and Mitigation of Stealthy Analog Hardware Trojans" [2310.03994]
- "DEMIST: A deep-learning-based task-specific denoising approach for myocardial perfusion SPECT" [2306.04249]
- "DEMIST: Decoupled Multi-stream Latent Diffusion for Quantitative Myelin Map Synthesis" [2511.12396]

Source: https://www.emergentmind.com/topics/demist