---
title: 'iThermTroj: Intermittent Thermal Trojan Attack'
url: https://www.emergentmind.com/topics/intermittent-thermal-trojans-ithermtroj
type: topic
---

# iThermTroj: Intermittent Thermal Trojan Attack

Intermittent Thermal Trojans (iThermTroj) are a class of thermal hardware Trojan attack for multi-processor System-on-Chips (SoCs) in which thermal sensor data are tampered with sporadically rather than continuously. In the defining formulation, the attack exploits chip thermal information in a random time-triggered manner: temperature manipulation is injected only at selected times, while the remaining readings are left untouched. The attack can lower, elevate, or fluctuate a victim core’s reported temperature, thereby misleading Dynamic Thermal Management (DTM) and degrading security, reliability, and performance while remaining more evasive than persistent thermal Trojans [2507.05576].

## 1. Definition and formal attack behavior

The defining characteristic of iThermTroj is intermittency. Conventional thermal Trojans are described as applying persistent, continuous temperature manipulation once active. By contrast, iThermTroj alter only a subset of sensor samples, with activation governed probabilistically by a chosen attack rate. The attack scenarios are explicitly given as Lowering, Elevation, or Fluctuation, with the attack rate selected from \(\{100\%, 80\%, 60\%, 40\%\}\), and a victim core chosen at random:

```latex
Attack_{Scenario} = Lowering, Elevation, or Fluctuation
Attack\_Rate = Pick \, from \, \{100\%, 80\%, 60\%, 40\%\}
IDX = Randomly \, Chosen \, Core \, Index
```

When the attack is applied to a steady-state thermal value, the reported temperature of the selected core is altered by an error term \(T_{error}\):

```latex
T(IDX) = T(IDX) - T_{error} \quad \text{(Lowering)}
```

```latex
T(IDX) = T(IDX) + T_{error} \quad \text{(Elevation)}
```

```latex
T(IDX) = T(IDX) \mp T_{error} \quad \text{(Fluctuation)}
```

The manipulated information consists of SoC thermal traces, steady-state temperature readings, and reported per-core temperature values used by thermal management logic. The attacker may use layout information together with the HotSpot 6.0 thermal simulator to generate thermal traces and then alter the victim core’s reported temperature. In operational terms, the attack flow is described as generating or obtaining thermal traces, selecting a victim core, injecting a temperature error \(\Delta t_{error}\), and repeating that process across only a portion of the data according to the attack rate.

This construction makes iThermTroj a data-manipulation attack on thermal telemetry rather than an always-on thermal distortion mechanism. A plausible implication is that the attack surface is not limited to thermal sensors themselves, but extends to any component that consumes or relays the thermal trace.

## 2. Distinction from persistent thermal Trojans

The paper’s central distinction is between persistent and intermittent thermal corruption. Persistent thermal Trojans continuously bias thermal readings after activation and therefore produce a stable abnormal signature. iThermTroj are instead intermittent, random time-triggered, and partial-data manipulators: they may affect only \(80\%\), \(60\%\), or \(40\%\) of thermal data rather than all samples [2507.05576].

The paper attributes the increased evasiveness of iThermTroj to three properties. First, intermittent manipulation does not create a constant abnormal thermal signature. Second, sporadic corruption can resemble ordinary sensor irregularities or noise. Third, static or threshold-based detectors are less effective because they are better matched to sustained anomalies than to sparse, stochastic deviations. The result is a threat model in which the thermal trace remains largely legitimate-looking even though DTM decisions are being perturbed.

This distinction also clarifies a common misconception: thermal Trojan behavior need not be continuous to be consequential. The paper’s main security message is precisely that mobile and multi-processor SoCs remain vulnerable even when only part of the thermal trace is modified.

## 3. Threat model, attack surface, and intended effects

The threat model assumes adversaries capable of compromising DTM in mobile SoCs through Hardware Trojan insertion or exploitation. The paper locates such adversaries across the semiconductor supply chain, including malicious insiders in design or fabrication, attackers exploiting untrusted third-party IP cores, and attackers using compromised EDA tools. Trojan placement is described near CPU or GPU cores, power management units, and analog or digital interfaces of thermal sensors.

The immediate objective is to manipulate thermal readings so that DTM makes incorrect control decisions. Under-reporting temperature can disable throttling, whereas over-reporting can trigger unnecessary throttling or reductions in clock frequency. The consequences described are not restricted to performance degradation; they also include reduced reliability, accelerated aging, and possible security disruption.

This threat model fits a broader Trojan literature in which malicious modifications remain dormant or low-observable until specific conditions arise. A survey of Trojans in digital manufacturing defines Trojans as “malicious modifications that may trigger their payloads at later times or stages of the product lifecycle,” and explicitly includes temperature among analog trigger sources; it also notes that Trojan activation is rare, which complicates detection [2301.10336]. Although that survey addresses Digital Manufacturing rather than mobile SoCs, it provides a compatible trigger-and-payload framework for understanding why thermal intermittency is operationally useful to an attacker.

## 4. Activation rates, duration scenarios, and stealth properties

The evaluated duration scenarios are \(100\%\), \(80\%\), \(60\%\), and \(40\%\), representing the fraction of thermal data points manipulated by the attacker. Even at \(100\%\), the attack is framed through random time-triggered application over the trace; at lower rates, it becomes explicitly partial, bursty, and sparse. The paper therefore does not model iThermTroj as an “always-on” perturbation, but as a randomized manipulation process over steady-state thermal samples.

The stealth advantage is attributed to the combination of partial corruption and non-periodic activation. Because only some data are modified, the thermal trace still largely looks legitimate. Because the corruption is not continuous or periodic, it does not expose a stable signature. Because the timing is random, it can mimic natural variation and sensor noise. The paper explicitly states that iThermTroj can bypass the most recent detection methods, including the Blind Identification Countermeasure (BIC).

A useful historical analogue appears in temperature-triggered Trojan work on ring-oscillator-based TRNGs. That attack remains dormant at ordinary temperature, introduces no additional logic, and becomes malicious only at elevated temperature, where entropy collapses and output predictability increases. It is not the same attack class as iThermTroj, but it provides a close precursor in which thermal conditions function as a stealthy, condition-based trigger [1910.00735]. This suggests that thermal intermittency can arise either from explicit trace manipulation, as in iThermTroj, or from latent physical conditions that selectively expose malicious behavior.

## 5. Failure of threshold-based detection and BIC limitations

The paper directly evaluates the inadequacy of threshold-based detection under intermittent conditions. It reports that BIC performs reasonably against a normal persistent thermal Trojan but loses effectiveness substantially against iThermTroj. With a conventional persistent attack, BIC maintains robustness for \(\Delta t_{error} > 3\) and \(\Delta t_{error} < -2\). Under iThermTroj, detection failures increase substantially across the same range, indicating that threshold-based behavior is not sufficient when anomalies are intermittent [2507.05576].

The paper also reports that BIC has a detection failure rate of \(52.38\%\) under certain conditions. More generally, existing methods are said to fail on smaller temperature modifications, reinforcing the vulnerability of simple threshold strategies. The important methodological point is that the detector is being defeated not by large-amplitude distortion alone but by the temporal structure of the corruption. Persistent anomalies are easier to isolate; sparse anomalies blend into background variation.

This observation resonates with earlier Trojan work grounded in semiconductor physics. Reliability-oriented hardware Trojans introduced through process deviation can remain functionally benign at first and then emerge later through temperature-accelerated degradation, time-based wearout, or condition-based triggers. That literature does not define iThermTroj, but it establishes a foundational principle: thermal conditions can make latent malicious behavior difficult to expose with ordinary production-time testing or static checking [0906.3834].

## 6. Tiny machine-learning countermeasures and measured results

To defend against intermittent thermal manipulation, the paper proposes tiny machine-learning classifiers for run-time anomaly detection in resource-constrained mobile SoCs. The rationale is deployment feasibility: the defense must be lightweight, low-latency, deployable on-chip, and able to monitor in real time. The evaluated classifiers are SVM, Logistic Regression, Random Forest, Decision Tree, and Naive Bayes in two variants. They are trained on steady-state temperature readings, including attacked samples, using a \(70\%\) training and \(30\%\) inference/testing split.

The evaluation uses accuracy, recall, precision, and F1-score, together with detection failure rate, resolution with respect to \(\Delta t_{error}\), and robustness under varying Trojan injection ratios. Compared to existing methods, the proposed approach improves attack detection rate by \(29.4\%\), \(17.2\%\), and \(14.3\%\) in scenarios where iThermTroj manipulates up to \(80\%\), \(60\%\), and \(40\%\) of SoC thermal data, respectively [2507.05576].

The reported ML-based detection failure rates are \(23\%\) at \(80\%\) Trojan injection, \(35.16\%\) at \(60\%\) injection, and \(38.10\%\) at \(40\%\) injection. The paper further reports a full protection resolution of \(0.8^\circ\text{C}\): any temperature manipulation exceeding \(\pm 0.8^\circ\text{C}\) is detected with \(100\%\) accuracy. It also specifies protection ranges by injection rate. At \(40\%\) injection, full detection is obtained for \(\Delta t_{error} < -2\) and \(> 2\); at \(60\%\) injection, for \(\Delta t_{error} < -1.2\) and \(> 1.2\); and at \(80\%\) injection, for \(\Delta t_{error} < -1\) and \(> 1\).

| Injection ratio | Detection-rate improvement | Full-detection range |
|---|---:|---|
| \(80\%\) | \(29.4\%\) | \(\Delta t_{error} < -1\) and \(> 1\) |
| \(60\%\) | \(17.2\%\) | \(\Delta t_{error} < -1.2\) and \(> 1.2\) |
| \(40\%\) | \(14.3\%\) | \(\Delta t_{error} < -2\) and \(> 2\) |

These results do not imply perfect security at low injection ratios; the nonzero failure rates show that intermittent attacks remain difficult even for the proposed defense. The paper’s strongest claim is therefore narrower: lightweight on-chip ML substantially improves runtime detection and extends the range of manipulations that can be fully detected.

## 7. Security significance, limitations, and research context

The principal implication is that thermal sensor integrity must be treated as a first-class security problem in mobile and multi-processor SoCs. DTM can be deceived by sporadic, randomized, partial manipulation of thermal traces, and threshold-based defenses are inadequate against that operating regime. The work therefore shifts the emphasis from persistent-abnormality detection to anomaly models that explicitly accommodate stochastic temporal sparsity [2507.05576].

Several limitations and open challenges are identified or implied. Detection is not perfect at lower injection rates, as reflected in the reported \(40\%\) and \(60\%\) failure rates. The study focuses on steady-state thermal data, leaving broader generalization to more dynamic workloads as an open question. Practical deployment of TinyML must still satisfy compute, memory, and latency constraints. The paper also notes the prospect of adaptive attackers that may use machine learning or analog circuitry to mimic benign noise and evade static countermeasures. Finally, because the experiments use a heterogeneous 6-core layout, extension across other SoC designs may require retraining or adaptation.

In the broader literature, iThermTroj sits at the intersection of thermal side channels, triggered payloads, and stealthy hardware Trojans. The digital manufacturing survey shows how temperature can serve as a trigger, how thermal behavior can support covert signaling or information leakage, and why rare activation complicates detection [2301.10336]. Earlier temperature-triggered Trojan work on TRNGs demonstrates that thermal conditions can selectively expose malicious behavior without altering nominal operation under standard conditions [1910.00735]. Reliability-oriented Trojan research further shows that temperature-accelerated physical mechanisms can be weaponized so that failure or degradation emerges only after stress or time [0906.3834]. Taken together, these works suggest that iThermTroj is best understood not as an isolated attack primitive, but as a specific SoC instantiation of a broader class of thermally mediated, low-observability Trojan behavior.

Source: https://www.emergentmind.com/topics/intermittent-thermal-trojans-ithermtroj