A new hardware Trojan detection technique using deep convolutional neural network
Introduction
Due to the huge requirements and the increasing complexities of the integrated circuit (IC) design, force the semiconductor industries to incorporate either third party Intellectual property cores (3PIP) or external vendors during the different phases of the IC supply chain. The involvement of these external parties may speed up the process and lessen the cost but also raises severe security concerns, including Hardware Trojan (HT), IP piracy, IC counterfeiting, IC overbuilding, and Reverse engineering [1]. Among all these threats, HT is the most dangerous due to its stealthy nature [2], [3]. Its consequences are very harmful and life-threatening for critical and sensitive applications such as Internet of Things (IoT) [4], [5], autonomous driving [6], surveillance systems [7], defense and mission-critical applications [8], [9]. An adversary can insert HT during the design and fabrication phase of the IC by either modifying the gate-level netlist or by manipulating the lithographic masks in the form of addition of gates, deletion of gates, and modification of size/parameters of gates (i.e., Parametric Trojans) [10], [11]. This malicious modification is triggered by rare condition and can change the functionality of the circuit, leak sensitive information, reduce the reliability, and cause a denial of service (DoS) [12].
Therefore, several countermeasures are reported in the literature that attempts to detect the Trojan during pre-silicon, and post-silicon IC phases [9], [13], [14]. However, in this paper, we are only concerned with the untrusted foundry problem, where an adversary can insert the Trojans during IC fabrication. Existing post-silicon techniques such as test pattern generation (TPG) and side-channel analyses (SCA) are either use several test patterns to activate the Trojans or measure various circuit parameters such as current, power, etc. for detection. However, TPG based approaches fail to detect large Trojans as it is difficult to generate a large number of test patterns to activate them. Whereas SCA based approaches are either rely on golden IC during detection or golden model-free, but fail to detect the small Trojans because they are vulnerable to process variations and fabrication noise [9], [13]. Hence, the aforementioned post-silicon techniques cannot cover all types of Trojans and possess a low detection rate. Recently several partial reverse engineering (RE) based techniques are also reported [15], which reduces the time and complexity involved in the full RE process [16], [17]. These techniques [18], [19], [20] extract the IC images and uses machine learning (ML)/image processing techniques to perform HT detection.
Bao et al. [18] proposed one-class support vector machine (SVM) based approach in which manual feature extraction takes place by comparing the IC layout image with golden IC image and five features are extracted to train the classifier for detection. Similarly, Abdurrahman et al. [21] proposed the Adaboost classifier-based approach in which they utilize the handcrafted feature selection algorithm HAAR to extract the features from IC schematic. However, both the manual and handcrafted features are not effective and highly affected by image variations. Moreover, manual feature extraction is a very complicated and time-consuming process, whereas handcrafted features are domain-dependent rule-based features selected by the experts, thus less effective for other domains. Further, they use shallow ML algorithms for detection which are less generalizable thus prone to underfitting/overfitting.
Therefore, to overcome the above-stated problems, we propose a robust partial RE based HT detection technique using Deep Convolutional Neural Network (DCNN). DCNN model have marked their success in the area of computer vision such as image classification [22], [23], object detection [24], [25], semantic segmentation [26] etc. It achieves state-of-the-art performance in multiple domains such as information security [27], remote sensing [23], medical field [28], [29] emotion recognition [30] and target detection [31] etc. by beating traditional handcrafted methods [32], [33]. The proposed DCNN model is different from ML/DL models because it contains the powerful feature extraction phase in its architecture, which uses the convolution operation rather than matrix multiplication. The inbuilt feature extraction process of the DCNN model is purely mathematical rather than domain-specific due to which it can extract the features from the images of any domain without requiring a handcrafted algorithm or golden IC reference. The reason behind its remarkable performance is the use of several convolutional and pooling layers, which perform the local connectivity and down-sampling to extract the relevant, invariant and non-linear features automatically in a hierarchical manner. The convolutional layer performs the convolution operation between the image and different kernels by establishing the local connection with the image subpart, which results in various unique feature maps. Max-pooling layer then reduces the size of each feature map by extracting the maximum information from it and provides partial and translation invariance, which makes the extracted features robust to any variation. All the feature maps are fed to the classification phase, which uses fully connected NN to learn these features and finally detect the IC as Trojan free, or Trojan inserted. To summarize, we have proposed a new DCNN model based HT detection technique and algorithm in this paper that accurately detects Trojans from IC layout images generated by partial RE. The major contributions are given as follows:
- (1)
A new DCNN model architecture is proposed that automatically extracts the relevant features from the IC layout images, even in the presence of any image and noise variations.
- (2)
A new training stopping condition method is proposed which uses two proposed metrics namely Accuracy difference measure (ADM) and Loss difference measure LDM to avoid the over-training of DCNN model.
- (3)
Experimental evaluation on synthetic ISCAS and Trust-Hub benchmarks provides on an average 97.4% and 99% accuracy, respectively.
The rest of the paper is organized as follows: Section 2 gives literature review analysis of existing HT detection techniques. Section 3 explains our proposed HT detection technique which includes proposed model architecture, stopping condition method and HT detection algorithm. Experimental results and comparative analysis is presented in Section 4. Finally, Section 5 concludes the paper.
Section snippets
Literature review: Analysis
Researchers in academia and industry have proposed several HT detection techniques at both pre and post-silicon phases. A detailed taxonomy of HT and their detection is presented in [13], [34]. Since we are only concerned with HT inserted during fabrication, therefore, analysis of the existing post-silicon techniques are presented in this section. Existing post-silicon testing based techniques are mainly TPG and SCA based. A sustained test vector technique is proposed in [35], which repeats the
New partial RE based hardware Trojan detection technique using deep CNN
This section first discusses the attack model and problem statement followed by the proposed DCNN model architecture, stopping condition method, and HT detection algorithm.
Experimental results and analysis
This section presents the experimental setup followed by the results and comparative analysis of the proposed technique.
Conclusion
This paper proposed a new partial RE based HT detection technique that detects the HTs inserted in the IC layout images using the DCNN model. In our technique, first, a new DCNN model is proposed, which automatically extracts the relevant features from the IC images without requiring any feature selection algorithm separately. To identify the stopping condition of the training process of the DCNN model, a new method has been proposed, which monitors the performance of the model using proposed
CRediT authorship contribution statement
Richa Sharma: Conceptualization, Methodology, Software, Writing - original draft, Formal analysis. Vijaypal Singh Rathor: Data curation, Validation. G.K. Sharma: Investigation, Supervision. Manisha Pattanaik: Supervision, Resources.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (62)
- et al.
Detecting malicious inclusions in secure hardware: Challenges and solutions
- et al.
A survey of hardware trojan threat and defense
Integration
(2016) - et al.
Multi-grade brain tumor classification using deep CNN with extensive data augmentation
J. Comput. Sci.
(2019) - et al.
Hardware trojan detection through golden chip-free statistical side-channel fingerprinting
- et al.
A primer on hardware security: Models, methods, and metrics
Proc. IEEE
(2014) - et al.
Hardware trojan: Threats and emerging solutions
- et al.
Privacy and security in internet of things and wearable devices
IEEE Trans. Multi-Scale Comput. Syst.
(2015) - et al.
Protecting endpoint devices in IoT supply chain
- et al.
Cybersecurity issues in robotics
- et al.
Safety and security in cyber-physical systems and internet-of-things systems
Proc. IEEE
(2017)
The hunt for the kill switch
IEEE Spectrum
Hardware trojan attacks: threat analysis and countermeasures
Proc. IEEE
Trustworthy hardware: Identifying and classifying hardware trojans
Computer
A survey of hardware trojan taxonomy and detection
IEEE Des. Test Comput.
Hardware trojans: Lessons learned after one decade of research
ACM Trans. Des. Autom. Electron. Syst.
Surveying the hardware trojan threat landscape for the internet-of-things
J. Hardw. Syst. Secur.
Hardware trust and assurance through reverse engineering: A survey and outlook from image analysis and machine learning perspectives
The state-of-the-art in IC reverse engineering
Steps toward automated deprocessing of integrated circuits
On application of one-class SVM to reverse engineering-based hardware trojan detection
Detecting hardware trojans using backside optical imaging of embedded watermarks
Detecting hardware trojans inserted by untrusted foundry using physical inspection and advanced image processing
J. Hardw. Syst. Secur.
Automatic feature selection of hardware layout: a step toward robust hardware trojan detection
J. Electron. Test.
Imagenet classification with deep convolutional neural networks
Deep feature extraction and classification of hyperspectral images based on convolutional neural networks
IEEE Trans. Geosci. Remote Sens.
Faster r-cnn: Towards real-time object detection with region proposal networks
A new two-level hierarchical diagnosis network based on convolutional neural network
IEEE Trans. Instrum. Meas.
Malware classification with deep convolutional neural networks
Medical image classification with convolutional neural network
EEG emotion recognition using dynamical graph convolutional neural networks
IEEE Trans. Affect. Comput.
Cited by (10)
WDP-BNN: Efficient wafer defect pattern classification via binarized neural network
2022, IntegrationCitation Excerpt :Thus, an automatic feature extractor and defect pattern classifier is needed. Convolutional neural network (CNN) has demonstrated great success in the semiconductor process in recent years [17–31]. However, CNN requires lots of computation due to the huge number of multiplications and additions, which makes it heavily rely on high-performance hardware [32].
Hardware Trojan Detection Using Shapley Ensemble Boosting
2023, Proceedings of the Asia and South Pacific Design Automation Conference, ASP-DACA Pre-Silicon Detection Based on Deep Learning Model for Hardware Trojans
2023, Journal of Circuits, Systems and ComputersA Hardware Trojan Detection and Diagnosis Method for Gate-Level Netlists Based on Different Machine Learning Algorithms
2022, Journal of Circuits, Systems and Computers