Hardware Trojan Detection With Shapley Boosting and Explainable Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning (ML) based detection methods for hardware trojans in integrated circuits face challenges such as lack of interpretability, inefficient feature selection, high training costs, and vulnerability to adversarial attacks, while side-channel analysis struggles with detecting subtle trojans due to process variation and noise.
Innovation Solution
Implementing Shapley Ensemble Boosting (SEB) to enhance ML models for hardware trojan detection, combining Shapley value analysis and boosting techniques to improve feature selection, model explainability, and robustness, using a synergistic integration of multiple lightweight models for parallel execution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning-based detection techniques are used for hardware trojan detection, then detection capability is provided, but the results are uninterpretable and feature selection is inefficient
Solution Approach 1:
The patent introduces SHAP (SHapley Additive exPlanations) values as an intermediary mechanism between the machine learning model and the detection results. SHAP values provide a mathematical framework to explain the output of the gradient boosting decision tree model by quantifying the contribution of each input feature to the prediction, thereby making the detection results interpretable while maintaining detection capability
Solution Approach 2:
The patent segments the feature selection process into distinct stages: initial feature extraction from circuit netlists, feature importance ranking using SHAP values, and selective feeding of top-ranked features to the gradient boosting model. This segmentation improves feature selection efficiency by filtering out irrelevant features before model processing
2Reliability
If traditional machine learning models are used for hardware trojan detection, then detection is performed, but training costs are excessive
Solution Approach 1:
The patent performs preliminary action by pre-ranking features using SHAP values before training the gradient boosting model. This preliminary feature ranking and selection step reduces the dimensionality of the input data, allowing the model to train faster with fewer parameters while maintaining or improving detection accuracy
Solution Approach 2:
The patent changes the parameter configuration by using a gradient boosting decision tree model with specific hyperparameters (max depth of 5, 100 estimators) optimized for hardware trojan detection. The model uses a fixed random state for reproducibility and processes features through a standardized pipeline, reducing training variability and computational overhead
3Reliability
If machine learning models are used for hardware trojan detection, then detection is provided, but the models are vulnerable to adversarial attacks
Solution Approach 1:
The patent applies preliminary anti-action by implementing adversarial training where adversarial examples are generated and used to retrain the gradient boosting model. This process strengthens the model's robustness against adversarial attacks by exposing it to perturbed inputs during training, enabling it to maintain detection accuracy even when faced with maliciously crafted circuit modifications
4Difficulty of detecting and measuring
If side-channel analysis is used for hardware trojan detection, then detection is attempted, but tiny hardware trojans hidden by process variation and environmental noise cannot be detected
Solution Approach 1:
The patent substitutes side-channel analysis (which relies on physical measurements susceptible to noise) with a netlist-based machine learning approach. The gradient boosting model analyzes structural features directly from circuit netlists, eliminating the impact of process variation and environmental noise that plague side-channel methods, thereby enabling precise detection of tiny hardware trojans
Data Source
AI summary
Various embodiments of the present disclosure provide hardware trojan detection using Shapley ensemble boosting. In one example, an embodiment provides for extracting a plurality of features related to hardware trojan detection from one or more circuit samples related to one or more circuits, training one or more machine learning models based at least in part on the plurality of features, modifying the one or more machine learning models based at least in part on a set of Shapley values to generate one or more enhanced machine learning models for hardware trojan detection related to the one or more circuits, and deploying the one or more enhanced machine learning models for the hardware trojan detection related to the one or more circuits.