1. Entropy Extraction in Metastability-based TRNG
Presented by
Cheng Chung Wang & Hsi Shou Wu

2. Outline
Background
Motivation
Implementation and analysis
Conclusion
Discussion

3. Background TRNG : True Random Number Generators
Entropy : Natural Sources
Cosmic rays
Stray electromagnetics waves
Thermal noise

4. TRNG Background cont’d Process variation and operating condition
will impact the output of TRNG circuits
Temperature 1
Fabrication defect
TRNG
Operating voltage

5. Background cont’d Bias Removal Techniques Post-processing techniques
Calibration techniques

6. Background cont’d Post-processing techniques XOR
Von Neumann corrector (entropy extractor)

7. Background cont’d
Calibration
Phase of clock signal
Charge injection

8. Motivation Several proposed biased removal circuit.
Which one is the best solution?
“Action speaks louder than words”

9. Implementation and analysis
TRNG without correction
Advanced technology
Technology scale down.
Need post-processing or calibration to boost entropy

10. Implementation and analysis
With XOR
Depend on the max entropy one
1. A dip in entropy of one of the TRNG will make the output dependent entirely on the other TRNG

11. Implementation and analysis
With Von Neumann corrector
Enhance entropy a lot
Drawback:
1. not generate at const rate!
2. Effective bit rate decrease with technology scaling

12. Implementation and analysis
TRNG with calibration
Tune the driving current/delay…

13. Implementation and analysis
TRNG with calibration
Increase entropy but remain const output rate
1. A comparison of the different bias removal techniques is shown in the Fig. 12. As seen from the simulation results, the circuit calibration technique provides significantly better correction in entropy as compared to the XOR function, with increasing device mismatch and improvement comparable to the Von Neumann technique, but at a constant bit rate.
2. Table.2 shows the average energy/bit for the basic TRNG and the different bias removal techniques. Although the XOR function incurs a very small overhead in the form of energy, its inefficiency with increasing variability does not make it a suitable candidate for usage in encryption systems designed in DSM technologies. The Von Neumann corrector maintains the entropy very close to one but at a significant energy cost. With increase in device mismatch more number of TRNG bits is needed per effective random bit generated. Hence, the energy per bit increases.
Less energy overhead

14. Implementation and analysis
TRNG with calibration
Tradeoff between number of bits entropy and energy consumption
1. Use less bit when no need for such high entropy and to decrease energy consumption

15. Conclusion
1. Modern security systems need on-chip true random number generators.
2. Conventional post-processing techniques are not efficient for simple TRNG
Physical calibration techniques are required .
provide a greater flexibility for trading off entropy for energy
1. A comparison of the different bias removal techniques is shown in the Fig. 12. As seen from the simulation results, the circuit calibration technique provides significantly better correction in entropy as compared to the XOR function, with increasing device mismatch and improvement comparable to the Von Neumann technique, but at a constant bit rate.
2. Table.2 shows the average energy/bit for the basic TRNG and the different bias removal techniques. Although the XOR function incurs a very small overhead in the form of energy, its inefficiency with increasing variability does not make it a suitable candidate for usage in encryption systems designed in DSM technologies. The Von Neumann corrector maintains the entropy very close to one but at a significant energy cost. With increase in device mismatch more number of TRNG bits is needed per effective random bit generated. Hence, the energy per bit increases.

16. Discussion Is the referenced model (dual inverter) representative?
Will the results change if we also run Monte Carlo simulation in other parameters (temperature or voltage drop)?
Would the area overhead be a huge issue?
As technology scaled down, dose the experiment result still make sense?
Is bit generation rate a more important issue?