TRUDEVICE 2013

Noise Reduction on Memory-based PUFs

Vincent van der Leest Roel Maes Geert-Jan Schrijen

Intrinsic-ID B.V.High Tech Campus 9

5656 AE Eindhoven, The Netherlands

Mafalda Cortez Said Hamdioui

Delft University of TechnologyMicroelectronics & Computer Engineering Lab

Mekelweg 4, 2628 CD DelftThe Netherlands

TRUDEVICE 2013Avignon, France, May 30-31, 2013

Extended version of this work will appear in HOST 2013M. Cortez, S. Hamdioui, V. vd Leest, R. Maes, G.-J. Scrijen, “Adapting Voltage Ramp-up Time for Temperature Noise Reduction on Memory-based PUFs”.

Noise Reduction on Memory-based PUFs

Security based on cryptographic algorithms Permanent key storage is highly

prone to physical attacks

Solution:1. Do not permanently store a key in

Non-Volatile Memories (NVMs)

2. Generate the key only when needed (extract it from a physical structure of the IC)

3. Delete the key

2

Motivation - Secure Key Storage

Physical Unclonable Functions(PUFs)

Device 1 Device 2

~ 50%difference

~ 10%errors

PUFs’ characteristics• Uniqueness

• Reproducibility

Noise Reduction on Memory-based PUFs

Motivation…………… shortcomings and contributions

3

• Shortcomings• Techniques to minimize the noise (increase reproducibility) of PUFs

• Our contributions• Efficient scheme to reduce the noise in memory-based PUFs

• Most of work focus on• Developing new PUFs types• Proving PUF uniqueness• Validating with silicon PUF reproducibility• Developing peripheral circuits (e.g., error correction)

Noise Reduction on Memory-based PUFs

Layout

4

• Memory-based PUF Secure Systems

• Enrolment versus reconstruction

• Stability parameters classification

• Technology versus non-technology parameters

• Simulation Set-Up & Results

• Industrial/Silicon results

• Block diagram of the new scheme

• Conclusion

Noise Reduction on Memory-based PUFs

Memory-based PUF Secure Systems

5

Enrollment: Define Key Reconstruction: Reconstruct Key

PRR and PR must be close enough!

PUF measurement

FuzzyExtractor

FuzzyExtractor

PUF measurement ’

Helper Data(public)

PRR(PUF Reference Response)

PR(PUF Response)

Key Key≈=

Reproducibility is crucial… but expensive!

Noise Reduction on Memory-based PUFs

Stability Parameters Classification

6

Stability parameters

Non-technologyparameters

Channel length

modulation

Threshold voltage

Temperature Supply voltage

Technology parameters

Geometry of transistors

LengthOxide

thicknessDoping

concentrationRamp-up

timeWidth

Manufacture Operation

NOTE: Classification based on: M. Cortez, A. Dargar, S. Hamdioui and G.-J. Schrijen, “Modeling SRAM start-up behavior for Physical Unclonable Functions”, DFT, 2012

Can non-technology parameters be used for noise reduction?

Noise Reduction on Memory-based PUFs

Simulation Set-Up & Results

7

…1 1 0 0 1 0

Monte Carlo

Vth1Vth2

Vth3Vthn Vthn+1

Vth1000

Pow

er-u

p

PUF fingerprint

Generation of an SRAM fingerprint of 1k

NOTE: Vth distribution according to: W. Zhao, F. Liu, K. Agarwal, D. Acharyya, S.R. Nassif, K.J. Nowka and Y. Cao, “Rigorous extraction of process variations for 65nm CMOS design”, ESSDERC, pp. 89–92, 2007.

Q5

Q1

Q6

Q2

VDD

Vin Vout

Set-up• HSPICE: SRAM cell with its periphericals • 45nm Low Power BSIM4 model

Noise Metric: Fractional Hamming Distance (FHD)

Experiments performed• Voltage Ramp-up Time : 3 tramp (10μs, 50μs and 90 μs)• Temperature : 3 Temp (-400C, 250C and 850C)• Measurements : 20 Meas (different noise random seeds)

Noise Reduction on Memory-based PUFs

Simulation Results… Noise at different enrollments

8

• For Temp below enrollment, max FHD is lower if tramp is longer • For Temp above enrollment, max FHD is lower is tramp is shorter

Can noise be reduced by manipulating Temp and tramp?

Typical approach is not optimal!

max FHD max FHD max FHD

Noise Reduction on Memory-based PUFs

Industrial/Silicon Results: Setup & Experiments performed

9

Measurment flow

Set temperature(-400C, +250C, +850C)

Read and store fingerprint

Power-down for 1 second

X9 (same Temp and Tramp)

x9

Power-up with tramp (10µs up to

500ms)

X 2

Experiments performed• Temperature : 3 Temp (-400C, 250C and 850C)• Voltage Ramp-up Time : 10 tramp (from 10μs up to 500ms)• Measurements : 10 Meas (each measurement has different noise)

Used devices

Noise Reduction on Memory-based PUFs

Industrial/Silicon Results: Original results

• Max noise 28%• Min noise 4.5%

10

Measured noise as FHD (precision 0.5%)

• Worse μ-BCHD = 0.37• Best μ-BCHD = 0.5

• Worse Hmin = 0.40• Best Hmin= 0.87

Optimization algorithms• Reproducibility

• Identify tramp for enrollment temp such that the reproducibility is the highest • Uniqueness

• Identify tramp for enrollement such that Entropy is the highest (Hmin)

Best value

Worse value

Legend

Noise Reduction on Memory-based PUFs

Industrial/Silicon Results: Reproducibility optimization algorithm

Main observations• Noise reduction for all devices at all conditions!• Fastest tramp during enrollment does not lead to the lowest noise• For Temp below enrollment, max FHD is lower for longer tramp

• For Temp above enrollment, max FHD is lower for shorter tramp

• Algortithm detorates μ-BCHD and Hmin for some devices (e.g., 130nm LP SRAM)

11

Noise Reduction on Memory-based PUFs

Industrial/Silicon Results: Reproducibility optimization algorithm

Main observations• Noise reduction for all devices at all conditions!• Fastest tramp during enrollment does not lead to the lowest noise• For Temp below enrollment, max FHD is lower for longer tramp

• For Temp above enrollment, max FHD is lower for shorter tramp

• Algortithm detorates μ-BCHD and Hmin for some devices (e.g., 130nm LP SRAM)

12

New best / worse• Max reduction 6% to 2% (3X)• New max noise = 17% (previous 28%)• New min noise = 2% (previous 4.5%)

Noise Reduction on Memory-based PUFs

Industrial/Silicon Results: Uniqueness optimization algorithm

13

Main observations• μ-BCHD and Hmin improved (or maintained for SRAM)• Fastest tramp during enrollment does not lead to the best Entropy (exception 130nm LP SRAM)

• For Temp below enrollment, max FHD is lower for longer tramp

• For Temp above enrollment, max FHD is lower for shorter tramp

• Noise increased in some cases (e.g., Buskeeper PUF)

Noise Reduction on Memory-based PUFs

Block diagram of the new scheme

14

Main observations• No adaptations of the PUF-circuit required• Standard components

Noise Reduction on Memory-based PUFs

Conclusion

15

• Novel idea for reducing noise on memory-based PUF fingerprints• Operational conditions• Temp versus tramp

• Validated using both SPICE simulation and silicon measurements

• Easy to implement

• Noise reduction achieved is up to 3x lower• Reduce overall cost and improve the robustness