💻 computer science

A Deep Cross Neural Network Framework for Stealthy Hardware Trojan Identification in FPGA

This paper proposes a Deep Cross Neural Network (DCNN) framework that achieves 98.4% accuracy in detecting stealthy Hardware Trojans in FPGA-based VLSI circuits by extracting and correlating structural and functional features, offering a reliable solution that reduces dependency on golden reference models.

Original authors: Kishor Bhagat, Gayatri Phade

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kishor Bhagat, Gayatri Phade

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive, incredibly complex Lego castle (an FPGA chip) that is supposed to do a specific job, like encrypting your bank data or guiding a satellite. Now, imagine a sneaky thief sneaks into the construction site and swaps out a few tiny, invisible Lego bricks for "spy bricks." These spy bricks look exactly like the real ones, but if a very specific, rare condition happens (like a secret code being typed), the spy bricks activate and cause the whole castle to malfunction or leak secrets. This is a Hardware Trojan.

The problem is that these spy bricks are so small and cleverly hidden that normal inspections (checking the blueprints or testing the castle under normal conditions) often miss them completely.

This paper proposes a new, super-smart way to find these spy bricks using a "digital detective" called a Deep Cross Neural Network (DCNN). Here is how the authors explain their method in simple terms:

1. The Problem with Old Methods

Traditional ways of finding these Trojans are like trying to find a needle in a haystack by looking at every single piece of hay one by one. They either need a "perfect" original blueprint (which we often don't have) or they try to measure tiny electrical vibrations (which is messy and easily confused by background noise). If the spy brick is very stealthy, these old methods often fail.

2. The New Solution: The "Group Hug" Detective

Instead of looking at the Lego bricks one by one, the authors' new system looks at how the bricks relate to each other.

Think of it like a classroom of students.

Old Method: The teacher asks each student, "Are you a spy?" One by one. A spy can easily say "No" and blend in.
New Method (DCNN): The teacher looks at the whole class and notices patterns. "Hey, usually when Student A raises their hand, Student B smiles. But in this group, when Student A raises their hand, Student B looks nervous." The teacher isn't just looking at the students; they are looking at the connections and relationships between them.

The authors call this "correlated learning." Their AI doesn't just check if a wire is too long or a gate is too big; it checks if the combination of features makes sense together. If a Trojan is hiding, it might change one tiny thing, but that change creates a weird ripple effect in how the other parts of the circuit behave together. The DCNN is trained to spot those weird ripples.

3. How the System Works (The 3 Steps)

The paper describes a three-step process to catch the spy:

Step A: Cleaning the Data (Preprocessing)
First, they take a huge library of circuit designs (some clean, some with Trojans) from a public database called "Trust-Hub." They clean up the messy data, removing the "noise" (like static on a radio) so the AI only sees the important signals.
Step B: Grouping the Clues (Clustering)
Before the AI tries to solve the mystery, it groups similar circuits together. Imagine sorting a pile of mixed-up socks. You put all the blue striped socks in one pile and the red polka-dot socks in another. This helps the AI understand that a "blue striped" circuit behaves differently than a "red polka-dot" one. By grouping them, the AI can focus on finding the one sock in the pile that doesn't quite match its group.
Step C: The Smart Guess (Prediction)
Finally, the Deep Cross Neural Network looks at these grouped piles. It uses its "cross" learning ability to see how different features talk to each other within a group. It asks, "Does this circuit's behavior match the pattern of a clean group, or does it show the subtle, weird connection pattern of a Trojan group?"

4. The Results: Catching the Sneaky Spies

The authors tested their system on standard benchmarks (the "Trust-Hub" library). Here is what they found:

Accuracy: Their system correctly identified the Trojans 98.4% of the time.
False Alarms: It rarely made mistakes, only flagging a clean circuit as bad 1.2% of the time.
Comparison: When they compared their "Group Hug" detective to other AI methods (like standard neural networks), their method was consistently better at spotting the stealthy, hard-to-find Trojans.

The Bottom Line

The paper claims that by teaching the computer to look at how different parts of a circuit work together (correlations) rather than just looking at them in isolation, we can catch malicious hardware modifications that were previously too clever to find. They didn't just build a better magnifying glass; they built a detective that understands the story the circuit is telling.

Note: The paper focuses strictly on detecting these Trojans in FPGA chips using this specific AI method. It does not claim to fix the Trojans, nor does it discuss using this for medical devices or other specific real-world applications beyond the testing environment described.

Technical Summary: A Deep Cross Neural Network Framework for Stealthy Hardware Trojan Identification in FPGA

Problem Statement
The increasing reliance on Field Programmable Gate Arrays (FPGAs) in security-critical applications (e.g., cryptography, aerospace, defense) has heightened vulnerability to Hardware Trojan (HT) attacks. These malicious circuits are intentionally inserted during design or fabrication to alter functionality, degrade performance, or leak data. Traditional detection methods face significant challenges:

Stealth: Trojans are designed with rare trigger conditions and minimal physical footprints, often evading functional testing.
Limitations of Existing Techniques: Functional testing requires impractical test patterns for large designs. Side-channel analysis (power, timing) is sensitive to process variations and environmental noise. Formal verification and reverse engineering often demand trusted "golden" reference models and high computational resources.
Machine Learning Gaps: While machine learning (ML) offers promise, existing classifiers often treat features independently, struggle with generalization across diverse architectures, and remain vulnerable to adversarial manipulation or data scarcity.

Methodology
The paper proposes a novel framework utilizing a Deep Cross Neural Network (DCNN) classifier to detect HTs in FPGA-based VLSI circuits. The approach is designed to minimize dependency on golden reference models by leveraging correlated feature learning. The workflow consists of three primary stages:

Data Preprocessing:
- Source: The study utilizes the Trust-Hub chip-level Trojan benchmark dataset, containing over 100 diverse designs (e.g., AES, RS232, microcontrollers) with both Trojan-free and Trojan-infected variants.
- Feature Extraction: Structural features (gate counts, fan-in/fan-out, signal path lengths) and functional/dynamic features (switching activity, toggle rates, power statistics) are extracted from netlists and simulation traces.
- Cleaning & Normalization: Data is cleaned to handle missing values and outliers. Features are normalized (min-max or z-score) and balanced using techniques like SMOTE to address class imbalances. The dataset is split into training, validation, and testing sets.
Clustering:
- Before classification, the processed feature set undergoes unsupervised clustering. This step groups circuit samples based on structural and functional similarity (e.g., gate complexity, connectivity patterns).
- Purpose: Clustering reduces intra-class variance, eliminates redundant data, and isolates hidden structures. It prepares the data by organizing samples into coherent groups, allowing the subsequent classifier to focus on correlated patterns rather than isolated anomalies.
Deep Cross Neural Network (DCNN) Classification:
- Architecture: The core of the framework is a DCNN designed to exploit correlations between multiple circuit parameters. Unlike standard classifiers that analyze features independently, the DCNN learns inter-feature relationships (e.g., simultaneous changes in gate count, fan-out distribution, and signal activity).
- Mechanism: The model uses input weight matrices ( $W$ ), recurrent weight matrices ( $U$ ), and bias vectors ( $b$ ). It employs a recursive composition mechanism where hidden states are updated based on cell states and connectivity filters. The training process minimizes Cross-Entropy loss via backpropagation to adjust weights.
- Prediction: The trained model maps clustered feature vectors to a correlation space, generating a binary classification (Trojan-free vs. Trojan-infected) based on the likelihood of Trojan presence.

Key Contributions

Correlated Neural Learning: The primary contribution is the application of a DCNN that explicitly models the correlation between structural and functional features, addressing the limitation of independent feature analysis in existing ML-based detection.
Clustered Feature Representation: The integration of a clustering pre-processing step to organize heterogeneous benchmark data, thereby enhancing the separability of Trojan-induced anomalies.
Reduced Dependency: The framework operates without requiring a trusted golden reference model or complex side-channel measurements, relying instead on the intrinsic structural and behavioral patterns of the circuit designs.
Scalability: The method is designed to be adaptable to complex FPGA architectures and diverse Trojan types (combinational, sequential, parametric).

Results
The framework was evaluated using the Trust-Hub dataset with varying training-to-testing ratios (60%, 70%, 80%). Performance was benchmarked against existing methods, including Knowledge Graph-based approaches (VD-KGE [24]) and Delay Side-Channel methods (ZDD [25]), as well as standard ANN and CNN models.

Accuracy: The proposed DCNN framework achieved a detection accuracy of 98.4%.
False Alarm Rate: The system demonstrated a reduced false alarm rate of 1.2%, a significant improvement over standard configurations.
Comparative Performance: The proposed method consistently outperformed the compared models (VD-KGE, ZDD, RNN, CNN) across all metrics, including Precision, Recall, F-Measure, Sensitivity, and Specificity.
Robustness: The model maintained stable performance across different Trojan types and circuit sizes. The Receiver Operating Characteristic (ROC) curve indicated an Area Under the Curve (AUC) approaching unity, signifying high sensitivity to stealthy Trojans while suppressing false positives.
Efficiency: The approach is noted for having "Low" computational overhead compared to graph-processing or trace-analysis methods, attributed to the efficiency of the clustered learning mechanism.

Significance and Claims
The authors claim that this work provides a reliable and efficient solution for enhancing security in FPGA-based VLSI systems. The significance of the study lies in its ability to identify subtle variations introduced by Hardware Trojans through correlated neural learning, overcoming the limitations of traditional testing and isolated feature analysis.

The paper asserts that the proposed algorithm effectively addresses key challenges in HT detection, including:

The stealthy nature of Trojan activation.
The need for scalability across diverse circuit architectures.
The requirement for robustness against adaptive attacks.

By reducing dependency on reference designs and achieving high accuracy with low false alarms, the framework is presented as a viable solution for practical hardware security applications. The authors suggest future work could focus on real-time detection, adaptive learning for evolving attacks, and optimization for low-resource FPGA environments, but the current study establishes the efficacy of the clustered DCNN approach on standard benchmarks.

1. The Problem with Old Methods

2. The New Solution: The "Group Hug" Detective

3. How the System Works (The 3 Steps)

4. The Results: Catching the Sneaky Spies

The Bottom Line

More like this