ML-Based Hardware Trojan Detection in ICs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting and mitigating hardware trojans in electronic systems are inefficient, requiring numerous binary monitors for each potential Trojan, leading to significant overhead and schedule impacts during the design and production of semiconductor devices.
Innovation Solution
A method utilizing machine learning instruments to capture and analyze operational data, develop normal and infected signatures, and detect anomalous behavior, allowing for the efficient selection and placement of detection monitors to identify and classify Trojans, thereby reducing the need for multiple monitors and minimizing physical and timing costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If numerous binary monitors are used for each potential Trojan, then detection capability is improved, but device complexity and overhead increase significantly
Solution Approach 1:
The patent applies universality by creating a single monitor that can detect multiple different Trojan types through machine learning analysis. Instead of having separate binary monitors for each potential Trojan, the system uses one monitor equipped with ML instruments that can identify various Trojan behaviors and patterns, thereby reducing the total number of monitors needed while maintaining comprehensive detection capability
Solution Approach 2:
The patent introduces machine learning instruments as an intermediary between the monitor and the Trojan detection process. These ML instruments analyze the operational data collected by the monitor and enable the system to identify different Trojan types without requiring specialized detection logic for each one, thus reducing device complexity while improving reliability
2Reliability
If numerous binary monitors are deployed for Trojan detection, then detection coverage is improved, but area, power, and timing costs increase
Solution Approach 1:
The patent reduces power consumption by replacing multiple specialized binary monitors with a single multi-functional monitor that uses machine learning to detect various Trojan types. This consolidation significantly reduces the total power required for monitoring operations while maintaining comprehensive detection coverage
Solution Approach 2:
The patent changes the operational parameters of the monitoring system by transitioning from fixed binary detection logic to adaptive machine learning-based detection. This parameter change allows the system to achieve comprehensive detection coverage with fewer resources, thereby reducing area, power, and timing costs
3Measurement precision
If traditional binary monitor methods are used, then detection of specific Trojans is achieved, but manual analysis and evaluation time increase
Solution Approach 1:
The patent implements self-service by enabling the monitoring system to automatically analyze and evaluate operational data using machine learning instruments. The system autonomously identifies Trojan patterns and classifies threats without requiring manual analysis, thereby maintaining high detection accuracy while significantly reducing the time loss associated with human evaluation
Solution Approach 2:
The patent incorporates feedback mechanisms where the machine learning instruments continuously analyze operational data and refine their detection capabilities. This automated feedback loop enables precise Trojan identification while eliminating the need for time-consuming manual analysis and evaluation cycles
Data Source
AI summary
A method and system for analysis of a facility may include providing an emulation host system, generating a pristine circuit model on the emulation host system, inserting a first hardware trojan model, emulating operation of the golden circuit model, and emulating operation of the first hardware trojan model, and determine a set of machine-learning models, detecting the presence of an unknown trojan as a function of the set of machine learning models.