Exploring Machine Learning Attacks on Logic Locking
1
Professor Nuzzo’s research group analyzes different types of hardware security attacks and defense techniques with the objective to enable systematic design of robust and secure IPs. Analysis of the existing approaches help reveal the current weaknesses and is key to the development of improved defense techniques. My next step is to finish my senior year and then apply to the electrical engineering program at USC. My advice for any future student participating in this program is to ask as many questions as possible. Know what your objective is and make sure everything is as clear as possible. I would thank Professor Pierluigi Nuzzo and my Ph.D. Mentor Subhajit Dutta Chowdhury for answering all my questions and giving me this opportunity. In modern times, security has become a major concern for any computing systems due to the globalization of the integrated circuit (IC) supply chain. Hardware security protects Intellectual properties (IPs) from attacks through various defense techniques like logic locking, gate camouflaging. Logic locking secures a circuit by adding extra logic gates called key gates, which hide the true functionality of the design. Existing logic locking techniques are often not secure and in this work we explore how Machine learning can be used to attack logic locking. We have used a Multi-layer Perceptron (MLP) as the neural network. The MLP analyzes the sub-circuit surrounding a key gate to directly predict its key value (0 or 1). NN Architecture: input features: 395 3 hidden layers: 1000, 512, 256 neurons NN training: 100,000 locality vectors, extracted from 10 different benchmark circuits locked with RLL, were used for training. NN testing: Each locked benchmark was used for testing Random Logic Locking (RLL): In RLL, the key gates (XOR/XNOR) are inserted at random locations in the IP. We have developed a machine learning based attack on Random Logic Locking. For each key gate inserted in the circuit, a corresponding locality vector (LV) is extracted from the netlist. LV extraction is based on the Breadth First traversal algorithm for graphs. Introduction Machine Learning Based Attack Logic Locking Next Steps; Advice for Future SHINE Students Acknowledgements Objective & Impact of Professor’s Research Exploring Machine Learning Attacks on Logic Locking Nic Hornung, [email protected] Mountain View High School, Class of 2022 USC Viterbi Department of Electrical and Computer Engineering, SHINE 2021 I have never coded a machine learning program previously but always wanted to try it. This program will greatly help with any coding classes I take in the future. Unlocked original circuit Original circuit locked with 2 keys Dataset Creation Key gate sub circuit Extracted LV for Key gate(KG) Benchmark Accuracy c432 55% c499 59% c880 52% c1355 54% c3540 51% c7552 52% The result shows that ML based approach can perform better than random guess. The current dataset is small. The prediction accuracy can be further improved by increasing the dataset size. Skills Learned -Boolean algebra and logic -Basic digital circuit design -Python coding -Researching a new and unfamiliar topic -Basics of neural networks (MLP) and their implementation in pytorch How This Relates to Your STEM Coursework
Transcript of Exploring Machine Learning Attacks on Logic Locking
Professor Nuzzo’s research groupanalyzes different types of hardwaresecurity attacks and defense techniqueswith the objective to enable systematicdesign of robust and secure IPs.Analysis of the existing approacheshelp reveal the current weaknesses andis key to the development of improveddefense techniques.
My next step is to finish my senioryear and then apply to the electricalengineering program at USC. Myadvice for any future studentparticipating in this program is toask as many questions as possible.Know what your objective is andmake sure everything is as clear aspossible.
I would thank Professor PierluigiNuzzo and my Ph.D. MentorSubhajit Dutta Chowdhury foranswering all my questions andgiving me this opportunity.
In modern times, security has become amajor concern for any computing systemsdue to the globalization of the integratedcircuit (IC) supply chain.
Hardware security protects Intellectualproperties (IPs) from attacks throughvarious defense techniques like logiclocking, gate camouflaging.
Logic locking secures a circuit by addingextra logic gates called key gates, whichhide the true functionality of the design.
Existing logic locking techniques are oftennot secure and in this work we explorehow Machine learning can be used toattack logic locking.
We have used a Multi-layer Perceptron(MLP) as the neural network. The MLPanalyzes the sub-circuit surrounding akey gate to directly predict its key value(0 or 1).
NN Architecture:input features: 3953 hidden layers: 1000, 512, 256 neurons
NN training: 100,000 locality vectors,extracted from 10 different benchmarkcircuits locked with RLL, were used fortraining.
NN testing: Each locked benchmark wasused for testing
Random Logic Locking (RLL):In RLL, the key gates (XOR/XNOR) areinserted at random locations in the IP.
We have developed a machine learningbased attack on Random Logic Locking.
For each key gate inserted in the circuit, acorresponding locality vector (LV) isextracted from the netlist.
LV extraction is based on the Breadth Firsttraversal algorithm for graphs.
Introduction Machine Learning Based AttackLogic Locking
Next Steps; Advice for Future SHINE
Students
Acknowledgements
Objective & Impact of Professor’s Research
Exploring Machine Learning Attacks on Logic LockingNic Hornung, [email protected]
Mountain View High School, Class of 2022USC Viterbi Department of Electrical and Computer Engineering, SHINE 2021
I have never coded a machinelearning program previously butalways wanted to try it. Thisprogram will greatly help with anycoding classes I take in the future.
Unlocked original
circuit
Original circuit locked with 2 keys
Dataset Creation
Key gate sub circuit
Extracted LV for Key
gate(KG)
Benchmark Accuracy
c432 55%
c499 59%
c880 52%
c1355 54%
c3540 51%
c7552 52%
The resultshows thatML basedapproachcan performbetter thanrandomguess.
The current dataset is small. The predictionaccuracy can be further improved byincreasing the dataset size.
Skills Learned
-Boolean algebra and logic-Basic digital circuit design-Python coding-Researching a new and unfamiliartopic-Basics of neural networks (MLP)and their implementation in pytorch
How This Relates to Your STEM Coursework
