Introduction
In the contemporary globalized semiconductor industry, integrated circuit (IC) design and manufacturing are frequently outsourced to third-party foundries and offshore vendors. While economically efficient, this complex supply chain introduces significant security vulnerabilities, most notably Hardware Trojans (HTs). These malicious modifications to a chip's logic can be designed to leak sensitive information, such as secret keys, cause catastrophic system failure, or provide unauthorized backdoors for attackers.
The primary challenge in hardware security is that Trojans are "needles in a haystack." They typically remain dormant, triggered only by extremely rare sequences of events, which makes them invisible to standard functional and structural testing. Traditional side-channel analysis, which monitors power consumption or heat dissipation, is often rendered ineffective by environmental noise and inherent manufacturing variations.
Methodology: Controlled Circuit Aging
To address these vulnerabilities, this research proposes a novel "active" detection technique called Controlled Circuit Aging. This approach leverages the physical phenomenon of Bias Temperature Instability (BTI), where transistors naturally slow down over time as their threshold voltage ($V_{th}$) increases.
The Synergy of Stress
The core innovation lies in the combined application of circuit aging and over-clocking. While neither factor alone is sufficient to reliably expose a Trojan, their synergy is powerful. By running an aged circuit at a clock speed exceeding its design specifications, specific timing violations are induced. Because a Trojan—no matter how small—introduces additional gates and wiring, it inevitably alters the timing characteristics of the paths it occupies. Aging magnifies these subtle delays, and over-clocking converts them into measurable bit errors at the output.
Feature Extraction and Machine Learning
The system monitors these timing violations, which manifest as bit flips. Four specific features are tracked: the number of $0 \to 1$ flips, $1 \to 0$ flips, and the weighted bit-locations of these errors. Detection is then performed using a two-stage machine learning pipeline:
- Autoencoder: Compresses the complex tensor of bit-error patterns into a compact, low-dimensional representation.
- One-Class Support Vector Machine (SVM): Learns the unique "fingerprint" of a clean, "golden" IC. Any deviation from this baseline is flagged as a potential Trojan.
Experimental Evaluation
The methodology was evaluated using industry-standard benchmarks from the Trust-Hub repository:
- BasicRSA-T100: A 32-bit RSA circuit where the Trojan is designed to leak private keys.
- AES-T100 & AES-T1000: 128-bit Advanced Encryption Standard circuits. These are particularly challenging as they contain "side-channel" Trojans that do not alter output bits under normal operating conditions.
The research utilized specialized standard cell libraries to simulate 45nm transistor behavior across various stages of degradation (duty cycles from 0% to 100%).
Key Results and Insights
The proposed method demonstrated exceptional performance across all benchmarks:
- High Accuracy: The system achieved a detection accuracy of $\\ge 99\%$.
- Extreme Sensitivity: It successfully identified a Trojan in the RSA circuit that occupied a mere 0.22% of the total circuit area.
- Detection of Stealthy Trojans: In the AES-T1000 benchmark—where the Trojan is located over 2,700 paths away from the critical path—the system maintained near-perfect precision by analyzing "batches" of inputs to refine the error signature.
- Robustness to Variation: The model accounts for real-world manufacturing variations (up to 17%), ensuring that natural chip-to-chip differences are not misidentified as malicious logic. Furthermore, the detection is input-independent; the ML model learns the statistical distribution of error propagation rather than specific input-output mappings.
Conclusion and Future Directions
Controlled circuit aging provides a robust, active mechanism for unmasking hidden hardware Trojans without requiring the activation of their malicious payloads. While highly effective in simulation, future work will focus on validating this approach on physical silicon to account for real-world environmental noise. Additionally, researchers are exploring the use of fast voltage switching to approximate aging effects more rapidly, potentially making this a practical "stress test" for high-volume manufacturing lines.
Summary updated